The sun · Greater Chicago Area
Hey, I'm Alex. I'm a cloud consultant at Spaulding Ridge and an Informatics graduate from Indiana University. Outside of work I build things: 3D worlds, simulators, a voice assistant, a macro-pad control deck, a city digital twin. This site is the tour.
Where I work and studied Cloud Consultant Spaulding Ridge Informatics graduate Indiana University What I'm working on now This website Building alextrosen.dev, the site you're on. ELARA Command Deck The pad arrived. Finishing live control on the real hardware. Chicago Alive, North Side Lincoln Park to Wrigleyville is built and in review, with the lakefront and the 'L' north to Sheridan. What I'm good at How these get made Most of what’s on this site was built with AI coding agents (Claude Code and Codex) working from my briefs. I decide what gets built, review it and test it. Each project keeps its tests and notes, and labels what is real and what is simulated.
On the workbench right now
A macro pad that approves my coding agents' requests. Chicago, one district tile at a time. A Godot rocket that still needs a human test pilot. This website. Say hi
Live · 2026 · last worked on Nov 7, 2026
The Constellation
A 3D galaxy where every project I've built is a world you can fly to.
The galaxy view from the showcase build (fixture data, nothing live). Every star system is a project.
The front door of the showcase build.
Type "take me to Data Atlas" and Jarvis flies you there.
A stop on Jarvis's tour, with captions and chapters.
Jarvis's tour stops on a walkable planet.
Use cases Alex opens the galaxy each morning to see the status of every project and launch the one he wants to work on. A first-time visitor takes Jarvis's narrated tour of the showcase build to see what Alex has built. Alex asks Jarvis by voice or text to fly him to a project's world and check its health. Land on The Constellation all 5 screenshots, the full story What it is A local-first 3D command center. Each star system is one of my projects, agents, notes or games, wired to real sources read-only, with anything simulated labelled as simulated.
Runs every day on my Mac mini for personal use. It is not a public app.
What it does
The Constellation is the place I keep everything. Open it and you’re looking at a galaxy: each star system is a project, a set of agents, a game or a collection. Fly into a system and you get that project’s world, with its real status pulled read-only from the source (git, the build pipeline, the app’s own health check).
It never pretends. If a world shows simulated activity, the screen says so.
How it’s built
A TypeScript server on Node 24 (Hono and the built-in node:sqlite) that only ever reads from the systems it watches.
A React 19 front end with React Three Fiber for the galaxy, systems and walkable planets.
Jarvis, the built-in guide, narrates a film-style tour and can fly the camera on request.
Vitest and Playwright suites gate every merge; nothing changes in the running copy until the tested tree is backed up and verified.
What’s next
More of the worlds get their own interactive builds, and the shareable showcase gets a proper home.
Highlights Galaxy, star-system and world camera levels you can travel between. Jarvis, a voice and hologram guide that gives a narrated tour of every world. Read-only adapters for git, coding-agent sessions, notes and a Roblox build pipeline. A health observatory and a one-click launcher for every project's real app. A static showcase build with a privacy check that scans every output file before it can ship. About 1,190 commits and 330+ unit-test files between Oct 18 and Nov 3, 2026. One hard problem The problem. A full-screen sky shader dragged an Intel Mac down to about 16 fps.
What fixed it. Baked the sky to a cube texture and ran bloom at half resolution, which brought it back to about 50 fps.
The source is private because it reads my own machine. A static, fixture-data showcase build exists for sharing.
Built with TypeScript React 19 React Three Fiber Three.js Hono node:sqlite Vite Vitest Playwright Private repository · 2026
Live · 2026 · last worked on Nov 2, 2026
Scalp City
A paper-trading research lab with a neon city for a dashboard, and no way to place a real order.
Use cases A researcher replays recorded market data through strategies that were pinned before they ran, in simulation only. Alex verifies the hash-chained market recorder to confirm the stored data has not been altered or corrupted. A reviewer reads the written incident records and the firewall rules that block every order path. Land on Scalp City all 3 screenshots, the full story What it is An isolated research platform that records market data into a tamper-evident store, runs pinned strategies against it and simulates fills in local ledgers. Paper only, by design.
Paper trading only. Live trading is locked and broker order paths are blocked in code. No results are published here.
What it does
Scalp City is a lab for testing short-term trading ideas without risking anything. It records market data, replays it through strategies that were pinned before they ran, and simulates what would have filled.
The city skyline is the dashboard: districts light up as the recorder, strategies and paper books do their work.
Safety first
The most important feature is the one that never happens. There is no live trading path:
The mode switch has a “live locked” state where every capability is off.
A firewall sits in front of the market-data client and refuses order and account endpoints outright.
Results stay inside the lab. You won’t find P&L, win rates or account figures anywhere on this site.
How it’s built
Python 3.12 with Starlette and httpx, SQLite for storage, and a hash chain over the recorded data so any tampering or corruption shows up on verify. Every incident gets a written record.
Highlights An append-only, hash-chained market recorder with an external checkpoint and a verify command. A firewall that checks every outbound request before any network I/O and denies all order and account paths. A credential-free, read-only local API with a strict content security policy. A paper engine with books, fees, risk limits and experiments. Strategies pinned by hash and pre-registered before they run, plus written incident records. One hard problem The problem. After a run of crashes, the market database came back corrupted. Identical bytes hashed differently on repeated reads, which pointed at faulty RAM, not the code.
What fixed it. Recovery stayed append-only. The store was cloned and recovered into a new database, the replacement was recorded as an explicit event, and the missing stretch was written down as a data gap instead of being papered over.
Private, and the dashboard stays private too. This page describes the architecture, not the numbers.
Built with Python 3.12 uv Starlette httpx websockets SQLite cryptography vanilla JS Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
ELARA Command Deck
A 12-key macro pad that lets me approve, reject or dictate to my coding agents from across the desk.
The deck map: every key and dial on one screen.
A pending approval, waiting for the second press.
High-risk requests get a warning and a confirmation on the Mac.
What the agent is doing right now.
The home screen at a glance.
Voice requests are reviewed on the Mac before they are sent.
Use cases A developer glances at a macro pad to see what Claude Code or Codex is asking, then presses twice to approve. Alex dictates a request to a coding agent with local Whisper and reviews the transcript on the Mac before sending. Alex gets a high-risk request such as a git push and has to confirm it at the Mac itself. Land on ELARA Command Deck all 6 screenshots, the full story What it is A local gateway that turns a NexMacro macro pad into a control surface for Claude Code and Codex: two-press approvals, voice dictation through local Whisper, and a little screen for every state.
Hooks, the Codex integration and transcription are verified live. Full pad control is still being finished now that the hardware has arrived.
What it does
Coding agents ask a lot of questions: can I run this, can I edit that. The Command Deck puts the answers on a physical pad next to the keyboard. One key approves, one rejects, one starts voice dictation, and the pad’s little screens show what’s being asked.
How it’s built
A Python gateway listens only on the local machine. A small Swift helper handles global hotkeys and audio. Voice goes through whisper.cpp on-device, and you review the transcript before it’s sent.
What’s next
Programming the pad over serial works. Next is finishing live control on the real hardware, then the microphone.
Highlights Loopback-only dashboard and JSON API with a host check and no CORS. A two-press approve flow bound to a request id, with a minimum display time so nothing gets approved by accident. High-risk requests can only be approved at the Mac itself. HMAC-signed Claude Code hooks, with Claude's own dialog as the fallback. Local voice dictation with whisper.cpp and a review step before anything is sent. A full set of state screens and icons drawn for the pad's display. One hard problem The problem. If an agent can read the gateway's files, could it approve its own requests?
What fixed it. After an independent review the approval path was rebuilt. The dashboard token lives only in memory and arrives by a one-time link. What's on disk can sign hook traffic or press harmless keys, but it can never approve.
Built with Python 3.12 Swift whisper.cpp Pillow Claude Code hooks Codex JSON-RPC Private repository · 2026
Working prototype · 2026 · last worked on Nov 4, 2026
Chicago Alive
A transparent 3D digital twin of Chicago's Loop, River North and Old Town, built from open data.
The downtown skyline view, with the layers panel and the replay timeline.
Willis Tower in the golden-hour light preset.
Down at the Chicago River, near Marina City.
Click any building to see where its shape and height came from.
The what-if model, clearly marked as hypothetical, with its uncertainty shown.
Old Town at night. Striped amber areas are honestly marked as not built yet.
Use cases A curious Chicagoan flies over the Loop and clicks a building to see where its height came from. A transit-minded visitor replays recorded 311, weather and CTA data on a timeline to see how a day unfolded. A planner asks what happens if a rail segment closes and reads the estimate with its P10 to P90 range. Land on Chicago Alive all 6 screenshots, the full story What it is A browser twin of part of Chicago, generated with a scripted Blender pipeline from open data. It replays recorded 311, weather and CTA data and runs a clearly labelled what-if model.
A working vertical slice covering about 2% of the city (63 district tiles, through Old Town). North Side builds 2–5, Lincoln Park to Wrigleyville, are built and waiting for review; they bring it to 147 tiles.
What it does
Chicago Alive is a slice of the city you can fly around in a browser. Every building comes from open data, and you can click one to see where its shape and height came from. Recorded city data plays back on a timeline, and a hypothetical “what if this rail segment closed” model shows its uncertainty instead of hiding it.
How it’s built
A headless Blender 4.5 pipeline generates the district tiles from source captures. The browser side is Three.js with MapLibre for the map, and a small Hono server serves recorded data. Areas that haven’t been built yet are marked on the map, in amber stripes.
What’s next
The North Side, in five planned builds.
Highlights 63 district tiles and 9 landmarks modelled from published dimensions. The CTA 'L' structure laid on real track geometry. A building inspector that shows each building's OpenStreetMap id and where its height came from. An offline replay timeline for recorded city data. A "close a rail segment" what-if model with P10–P90 bands, kept visually apart from observed data. Hash-pinned source captures, each with its licence. One hard problem The problem. The skyline view crawled at 10 fps with 3,177 draw calls and 21 million triangles.
What fixed it. Instanced train cars and merged, skinned pedestrians cut it to 694 draw calls and 7.7 million triangles, and the view now hits the browser's 60 fps frame cap.
Built with TypeScript Three.js MapLibre Blender 4.5 (scripted) Hono zod node:sqlite Vite Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
Orbital Foundry
Build a spacecraft, launch it at real Earth scale and fly it to the Moon with real orbital mechanics.
The hangar, where staging and delta-v update as you build.
On the pad, moments before liftoff.
In orbit at real Earth scale. The assists fly through the same controls you do.
Planning a lunar transfer on the mission map.
Rendezvous and docking in the Rescue Window mission.
After every flight, a review with the full telemetry.
Use cases A spaceflight fan builds a craft from original parts and checks delta-v and thrust-to-weight stage by stage. A player flies First Orbit, Lunar Courier or Rescue Window and plans burns with maneuver nodes. A student reviews a flight's telemetry graphs, exports the CSV or replays it deterministically. Land on Orbital Foundry all 6 screenshots, the full story What it is A native macOS aerospace sandbox in Godot. You design a craft in a hangar from original parts, launch, and fly three missions where orbits come from gravity and thrust, not scripts.
All three missions complete in automated tests. A first hands-on playtest is still to come, and the app isn't notarized for distribution.
What it does
Orbital Foundry is a spaceflight sandbox. You put a rocket together, launch it from a pad at real Earth scale, and fly missions that end in lunar orbit and back. There are no rails: if you get into orbit, it’s because the physics worked.
How it’s built
Pure GDScript on Godot 4.7, running on Metal. Assists are available, and they fly the craft through the same controls the player uses, so they can’t cheat.
What’s next
A proper hands-on playtest, then signing the app so other people can try it.
Highlights 25 original parts and a staging rail that shows delta-v and thrust-to-weight per stage. A patched-conic map with a maneuver-node editor and Moon encounters. Mission review with telemetry graphs, CSV export and deterministic replay. Three missions, First Orbit, Lunar Courier and Rescue Window. Physics checked against textbook values (Hohmann transfers, vis-viva, Kepler's laws). One hard problem The problem. Godot's 32-bit vectors can only resolve about 32 metres at the Moon's distance, so spacecraft jittered.
What fixed it. Positions live in double precision with a floating origin, and distant bodies are drawn in a scaled space. A test checks the precision at lunar distance.
Built with Private repository · 2026
Working prototype · 2026 · last worked on Nov 2, 2026
Project Meridian
An expedition game where the weather and earthquakes come from real recorded data.
The mission room. Recorded and synthetic days are labelled differently.
Every number can be traced back to its source.
Planning: pick a route and kit, and see simulated odds before you commit.
Out on the ice. Wind and visibility come from the recorded day.
The debrief. The server re-runs the expedition before the score counts.
Use cases A player picks a region, reads the real recorded weather and quake conditions and chooses a kit for the expedition. A curious player opens the provenance inspector to see each value's unit, request URL, hash and licence. A player replays a run with counterfactuals to learn how a different kit or route would have gone. Land on Project Meridian all 5 screenshots, the full story What it is An original browser game. Each expedition's conditions come from recorded Open-Meteo weather and USGS quake data, and the server re-simulates every run before it records a score.
Playable with 3 regions. Route choice still matters less than kit choice, and the AI adviser isn't built.
What it does
Pick a region, read the conditions, choose your kit and route, and see if the expedition makes it. The twist is that the conditions are real: recorded weather and earthquake data from public sources, frozen into hash-checked snapshots.
How it’s built
React and Three.js with a MapLibre globe on the front end; a Hono server with SQLite that re-runs every expedition before a score counts, so the client can’t fake one.
Highlights 3 regions and 9 hash-pinned recorded data bundles, plus clearly labelled synthetic days. A provenance inspector that shows every value with its unit, request URL, hash and licence. Replayable runs with counterfactuals. Live data older than 3 hours gets a STALE badge. Modes are never quietly upgraded. Built with TypeScript React Three.js MapLibre Hono zod node:sqlite Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
The Arena
A tournament where coding agents compete, and an independent judge checks the work instead of trusting the claim.
The starting page, showing the bundled example tournament.
One attempt, replayed event by event, with the evaluator's verdict.
The five tasks, with budgets and public rubrics.
The agents, including one built to overclaim so the judge can catch it.
A passing attempt, with its budgets and timeline.
Use cases A developer compares two coding agents on the same five seeded tasks and reads the scores from the evaluator. A researcher replays the whole tournament to confirm it gives the same scores, then reads the transcripts. A team uses the dishonest Overclaimer agent to confirm the evaluator catches false claims of success. Land on The Arena all 5 screenshots, the full story What it is A local, reproducible tournament facility for tool-using agents. An evaluator inspects the workspace after every attempt, and the whole tournament can be re-run to identical scores.
What it does
The Arena runs agents through the same tasks under the same conditions and scores what they actually did. Nothing an agent says about itself affects its score.
How it’s built
Node 24 with its permission model as the sandbox, a Hono API and SQLite event log, and a React viewer for replaying transcripts.
Highlights Five seeded tasks, from patching code to reconciling records to surviving a flaky service. A hash-chained, append-only event log with transcript replay. A re-run of the whole tournament gave identical scores on 20 of 20 attempts. A deliberately dishonest "Overclaimer" agent as a negative control. One hard problem The problem. Agents say they finished. That isn't the same as finishing.
What fixed it. A submission only counts as a claim. The evaluator checks the workspace itself, and timeouts, budget overruns and crashes score zero even if the files happen to be right.
Built with TypeScript Node 24 Hono node:sqlite React Vite Playwright Private repository · 2026
Live · 2026 · last worked on Oct 29, 2026
Last Legend
Draft up to 100 athletes, mascots and legends, close the storm, and see who's left standing.
The home page of the battle-royale simulation.
The draft screen, which says plainly when a roster feed is not loaded.
The library of selectable legends.
Champion history is stored on the device only.
Use cases A group of friends drafts a field of athletes, mascots and legends and watches the storm pick a champion. A fan replays the same seed to settle an argument, because the same seed gives the same story. Someone on a phone runs a 100-contestant match in a few seconds and shares the seed. Land on Last Legend all 4 screenshots, the full story What it is A static browser battle-royale simulator. Seeded and replayable, with procedural arenas and a storm that closes in until one champion remains.
Live on GitHub Pages. Ratings are made-up game numbers and balance is still a work in progress. It's a hobby simulator, with no affiliation to any athlete, team or league.
What it does
Pick a field, hit go, and watch a storm squeeze the arena until one name is left. Same seed, same story, every time.
How it’s built
No framework: plain HTML, CSS, SVG and JavaScript, with a Python reference engine to check the browser one against.
Highlights 1,200+ selectable entries and matches of 2 to 100 contestants. Seeded, deterministic replays. Four procedural arenas. A 100-contestant match simulates in about 0.6 seconds, and the layout holds up at 390px wide. Built with JavaScript HTML CSS SVG Python (reference engine) Public repository · 2026
Play it ↗
In progress · 2026 · last worked on Nov 4, 2026
DataAtlas
A museum of public-data experiences built from official U.S. and NASA datasets.
The museum's front page and its Discovery of the Day.
The index of all the exhibits.
The Data Time Machine: a century of state statistics on an animated map.
The Economic Galaxy places every large U.S. city as a star by its economics.
City DNA turns a city's census figures into a radial fingerprint.
The Ranking Race animates two centuries of city rankings.
Use cases A student opens an experience to learn what a public dataset measures, where it came from and what it shows. A teacher searches across the entities to find a dataset story that fits a lesson. A visitor checks the Discovery of the Day for one verified item drawn from official U.S. and NASA data. Land on DataAtlas all 6 screenshots, the full story What it is A planned set of 15 interactive data stories, built as a static site. Two experiences are finished; the rest are in progress or paused.
Paused. Two of fifteen experiences are done and it isn't deployed.
What it does
Each room in the museum is one dataset told well: what it measures, where it came from, and what it shows.
What’s next
It’s paused while other projects get attention. The two finished rooms are the template for the rest.
Highlights 22 chart building blocks and a chart recommender that refuses any chart that would hide rows. A polite data pipeline with caching and validated writes. A "Discovery of the Day" pool of 359 verified items. Search across about 11,500 entities. One hard problem The problem. Several government APIs started requiring keys.
What fixed it. The pipeline switched to keyless bulk downloads and public JSON endpoints, so the site never needs a secret.
Built with Astro D3 Canvas TypeScript Vitest Playwright axe Private repository · 2026
Live · 2026
Wobble Bay
A Roblox collecting game, made with an AI-assisted game pipeline I built for it.
A view of the live game.
Wobble Bay's world in my Constellation command center.
The same world from a low angle across the water.
A closer look at the island.
The project card shows its release state and test results.
Use cases A Roblox player wanders the bay collecting creatures of different rarities, with day, night and weather spawns. A player fills the collection by visiting different areas at different times of day. Alex uses the factory to build and test updates inside Roblox Studio before releasing them. Land on Wobble Bay all 5 screenshots, the full story What it is A creature-collecting Roblox game, and the first one out of my Roblox Game Factory, a local pipeline that builds and tests games with coding agents and Roblox Studio.
Published on Roblox. It's tiny so far, and the factory around it is a functional prototype.
What it does
Wander the bay, find wobbly creatures, fill the collection. Behind it is the factory: a dashboard and job runner that builds games with coding agents and tests them inside Roblox Studio before anything is released.
Highlights Collectible creatures in several rarities, spread across different areas with day, night and weather spawns. The factory's hash-chained audit log, verified backups and owner-only release gates. An automated Studio test harness driven through Roblox Studio's MCP server. Built with Luau Rojo Roblox Studio TypeScript Hono React node:sqlite Private repository · 2026
Live · 2026 · last worked on Oct 26, 2026
Cycle 96
A research and paper-trading platform with a custom dashboard; no real money.
Use cases A researcher tests a trading idea in simulation only, with no real orders possible, and records the plan first. Alex lets a background service run paper experiments unattended on his own computer and reviews them on a custom dashboard. A reviewer follows the staged approval ladder to see how an idea moves from a written hypothesis to a paper run. Land on Cycle 96 the full story What it is A local research platform that evaluates ideas in simulation only. No production order path exists in the code.
Paper only. It runs every day on my own machine; nothing is public and no results are shown.
Highlights Simulation only: no production order submission exists. Runs unattended as a background service on my own computer. Pre-registered research workflow with a staged approval ladder. Described in general terms only, on purpose. No screenshots, data or results appear on this site.
Built with Python FastAPI DuckDB pandas scikit-learn Automated test suite Private repository · 2026
Working prototype · 2026 · last worked on Nov 4, 2026
Linewise
A private sports data research and record-keeping web app.
The odds board compares books side by side; every price is blurred.
Game cards with spread, total and moneyline columns, numbers blurred.
Its world in my Constellation command center.
Use cases A sports data hobbyist keeps research notes and records in one invite-only app on their own machine. Alex tries the app with simulated demo data, with no outside keys needed and every screen labelled as simulated. A tester runs the desktop and phone end-to-end suites before trusting a change. Land on Linewise all 3 screenshots, the full story What it is An invite-only web application for sports data research and record-keeping. It runs on simulated demo data by default.
Runs locally on simulated demo data. Not deployed.
Highlights Runs entirely locally with no external keys needed. Unit and end-to-end test suites (desktop and phone). Every demo screen is labelled as simulated. Described in general terms only, on purpose. The screenshots show simulated demo data, with every price, figure, date and time blurred.
Built with Next.js React TypeScript Drizzle ORM Embedded Postgres (PGlite) Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
NexMacro for Mac
A native macOS toolkit (library, CLI and menu-bar app) for configuring a USB macro pad and driving its backlight.
padctl, the command-line tool: every command that changes the pad needs --yes.
Without --yes, padctl prints the exact frames it would send and touches nothing.
Applying a ready-made Mac layout, previewed key by key before anything is written.
Use cases A Mac user with a USB macro pad configures its keys from a command-line tool instead of the vendor app. Alex changes the pad's backlight from the menu bar so it reflects the state of his system. A cautious user backs up the pad's configuration before writing, with an explicit flag required to confirm. Land on NexMacro for Mac all 3 screenshots, the full story What it is A Swift package that speaks the macro pad's serial protocol, with a command-line tool and a menu-bar app that controls lighting and mirrors system status.
Highlights Library, CLI and menu-bar app in one package. Writes require an explicit confirmation flag; back up first. Menu-bar app running on my machine since 2026-11-02. Built with Swift 6 Swift Package Manager macOS menu-bar app Serial protocol Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
Multi-AI Command Center
A local control surface for Claude Code and fallback coding agents with usage gauges and verified handoffs.
The control surface: projects, chats, approvals and a usage meter.
The overview tab, with the recent-sessions list blurred.
Panels are tabs you can rearrange.
Use cases A developer who hits a usage limit in one coding agent hands the work to another through a verified checkpoint. Alex checks real usage gauges for his coding agents before deciding which one to start a task with. A user resumes or forks an earlier agent session from its transcript on disk, in one browser workspace. Land on Multi-AI Command Center all 3 screenshots, the full story What it is A browser workspace that drives real agent sessions, shows real usage and hands work between agents through verified checkpoints.
Highlights Streaming chats, approvals and a real terminal in one workspace. Usage gauges from real telemetry, never mock data. Resume or fork sessions from transcripts on disk. Built with TypeScript Node.js React Vite Tailwind xterm.js Monaco Claude Agent SDK Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
Claude Free Swarm
A free-first team of local and free-tier AI workers that Claude Code delegates to over MCP.
swarm doctor checks every provider, the kill switch and that spend stays at zero.
The model registry: which free or local models are usable, and where prompts go.
Agents talk to each other directly over the message bus, as this past run shows.
Its world in my Constellation command center.
Use cases A developer hands a small bug fix to local and free-tier workers and gets back a patch with test results. Alex asks several specialist agents a research question and receives one combined, reviewed answer instead of five separate replies. A cost-conscious user checks the budget tracker and uses the kill switch to stop a run that is going wrong. Land on Claude Free Swarm all 4 screenshots, the full story What it is An MCP server, CLI and skills that dispatch code, bug, research and council tasks to specialist agents that message each other and return one verified result.
Works for small tasks; the larger demo runs failed in the last benchmark.
Highlights Free-first by design; it never commits, pushes or deploys on its own. Coordinator-run tests and reviewer verdict on every patch. Kill switch and budget tracking built in. Built with Python MCP SQLite message bus Git worktrees Local LLM runtime Private repository · 2026
Working prototype · 2026 · last worked on Nov 2, 2026
Chronos
An evidence-linked historical 3D reconstruction explorer with a confidence level on every structure.
A fictional district shown as it stood in 1871, with a banner saying it is synthetic.
Every element is tagged documented, inferred or speculative, with a score.
The same district after the second survey, 1893, with new buildings.
The audit tab runs the dataset's acceptance checks and records each run.
Use cases A history teacher drags the time slider between two layers to show how a district changed. A student clicks a building to see its sources and whether it is documented, inferred or speculative. A researcher checks the dataset audit and hatched shading to separate what is known from what is guessed. Land on Chronos all 4 screenshots, the full story What it is Every building carries sources, rights credits and a documented/inferred/speculative rating; uncertain items are drawn hatched, never as solid fact. The demo district is fictional.
Highlights Time slider between two historical layers. Evidence panel per structure. Dataset audit run on start. Clearly bannered as synthetic demo data. Built with TypeScript React Three.js Vite Vitest Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
Mirror
Verifiable photos-to-3D reconstruction of a tabletop object that reports its own measurement error.
A synthetic test object reconstructed from photos, with its measured size compared to the known one.
The ingest tab: every photo is checked and its location data scrubbed.
Each pipeline stage with real timings, and the one that can't run here marked unsupported.
Exports come with hashes you can re-check in the browser.
Use cases A maker photographs a small tabletop object and builds a sparse 3D point cloud from the photos. A student recovers real scale from printed markers and compares the measured size to a value they typed in. A researcher checks the error readout to see how far the reconstruction is from a known measurement. Land on Mirror all 4 screenshots, the full story What it is Estimates camera poses, builds a sparse point cloud, recovers scale from printed markers and compares the measured size to a typed-in value; unsupported stages are labelled as such.
Highlights Full sample run in about a minute on CPU. Explicit error readout against a known measurement. Dense reconstruction honestly marked unsupported. Built with Python COLMAP TypeScript React Three.js Vitest Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
Vanguard
A safe simulated robotics academy: a warehouse robot with classical planning beside a learned policy.
A simulated warehouse robot and its planned route.
After an episode, the scorecard shows delivery, path length and collisions.
A benchmark compares policies on the same seeds.
Training imitates the classical planner, and the page says it is not reinforcement learning.
Use cases A robotics student compares a classical A* planner and a learned policy on the same warehouse layouts. A learner tries a lidar-equipped robot with a battery budget in simulation with no hardware path. A researcher runs the fixed-seed benchmark on held-out layouts and replays the result in the viewer. Land on Vanguard all 4 screenshots, the full story What it is A 2D kinematic robot simulator with lidar, a battery budget, A* planning, a Gym-style interface, held-out benchmarks and a replay viewer. Simulation only by design.
Highlights Fixed-seed benchmark with held-out layouts. Enforced no-hardware path (tested). 92 unit and integration tests. Built with Python FastAPI NumPy Playwright pytest Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
BRIDGE
Describe in plain English how an input device should control an action, review the diff, then arm it.
The workbench: keyboard, pad, wheel and gamepad inputs on the left, mappings and targets on the right.
A plain-English sentence becomes reviewable mappings that change nothing until applied.
The same idea aimed at a small on-page instrument instead of a whiteboard.
Use cases A teacher types that the space bar should advance slides and reviews the proposed mapping before arming it. A gamepad owner describes how a button should drive the built-in synth and checks the diff before applying it. Someone testing a new input device hits the stop button or Shift+Esc to halt every mapping at once. Land on BRIDGE all 3 screenshots, the full story What it is Maps keys, clicks, wheels, gamepads and HID inputs to its own built-in targets (whiteboard, slides, player, synth). A kill switch halts everything.
Highlights Deterministic, offline prompt parser with visible assumptions. Apply-only-on-approval diffs. Stop button and Shift+Esc halt all mappings. Refuses to control other apps. Built with TypeScript React Vite Node.js Vitest Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
FOLD
A sentence plus object dimensions become custom packaging, folded in 3D beside an exact-scale dieline.
A prompt becomes a box in 3D and as a flat pattern with fold lines.
The Unfold slider flattens the box so you can see how it is made.
Switching construction shows what changed in size, board and folds.
Use cases A packaging designer roughs out a tuck-end carton for a new product and prints the exact-scale dieline. A small shop owner describes a box around given object dimensions, then revises it in plain English. A print vendor receives a 1:1 SVG or PDF with separate cut and score layers. Land on FOLD all 3 screenshots, the full story What it is Proposes real carton constructions, lets you revise in plain English and exports 1:1 SVG and PDF dielines with separate cut and score layers.
Highlights Four constructions from one parameter model. Exact-scale SVG/PDF export with cut and score layers. Six acceptance demonstrations via self-test. Built with TypeScript React Three.js Node.js Vitest Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
SPARK
Describe how a low-voltage gadget should react; get a state machine, simulator, wiring plan and firmware source.
A sentence becomes a state machine and a simulated wiring bench.
A different sentence, a different circuit.
The event theater replays the behaviour against a scripted button, next to the parts list.
Use cases A hobbyist describes how a low-voltage gadget should react and tests it on a live virtual device. An electronics beginner gets a parts and power plan for a microcontroller before buying anything. A maker generates firmware source from the same spec that drove the simulator. Land on SPARK all 3 screenshots, the full story What it is One spec drives a live virtual device, a parts and power plan for a microcontroller, and generated firmware source.
Highlights Typed state machine with a live virtual device. Wiring and power budget output. Firmware source generation. Eight acceptance scenarios. Built with TypeScript React Vite Node.js Vitest Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
SYNAPSE Flow
Describe a workflow in a sentence; get a typed, inspectable, permission-scoped flow that dry-runs before it acts.
Describe a workflow in a sentence, or start from a template.
The sentence and the wiring it produced.
Circuit mode shows the workflow as nodes you can edit.
A dry run reports what would happen without writing or sending anything.
Use cases Someone who wants new files from a folder sent to a webhook describes it in a sentence and inspects the graph. A cautious user dry-runs a workflow to see the exact paths and outbound requests before it acts. A user revises a flow in plain English and checks the new revision before running it. Land on SYNAPSE Flow all 4 screenshots, the full story What it is Turns a sentence into a graph between folders, feeds and webhooks, shows example data moving along it, and only runs for real after a dry run.
Highlights Offline deterministic parser that lists assumptions. Dry-run lab shows exact paths and outbound requests. Revisions by plain-English edits. 29 unit tests plus end-to-end test. Built with TypeScript React Vite Node.js Vitest Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
FRANKENAPP
Describe features you wish existed together and get three working mini-app candidates over your own data.
Use cases A freelancer who tracks clients in a spreadsheet describes the features they want and picks one of three candidate mini-apps. Someone imports a CSV of their own records and reshapes the app in words without losing any rows. A user undoes a change, then exports the data back to CSV or JSON when finished. Land on FRANKENAPP the full story What it is Builds validated app specs from an allowlist of components (no generated code), lets you import CSV or JSON, and reshape the app in words without losing records.
Highlights Three structurally different candidates per prompt. Runtime built from a fixed gene allowlist. Undo/redo and CSV/JSON import and export. Built with TypeScript React Zod Vite Vitest Playwright Private repository · 2026
Working prototype · 2026 · last worked on Oct 9, 2026
DevForge
Free browser-only developer tools plus three premium code packages, all built to run at zero cost.
The tool index: everything runs locally in the browser.
The JSON formatter, with an example pasted in.
The regex tester highlights matches and explains each part.
The three paid code packages, with free samples.
Use cases A developer pastes a JSON payload into a browser-only formatter without sending it to any server. A backend engineer tests a regex or checks a cron schedule in a tool that makes no outside requests. A team evaluates a dependency-free premium package using its docs, examples and free sample before using it. Land on DevForge all 4 screenshots, the full story What it is Twelve privacy-respecting tools (JSON, regex, SQL, cron and more) and three dependency-free packages, with tests enforcing that tools make no outside requests.
Highlights 12 browser-only tools with a strict content security policy. About 1,150 automated tests across tools and packages. Packages ship with docs, examples and a free sample. Built with JavaScript Static site Web Components Playwright Node.js Private repository · 2026
Working prototype · 2026 · last worked on Nov 3, 2026
PuzzleForge
A generator of machine-verified printable puzzle packs, with a free promo site and sales dashboard.
The storefront for printable puzzle packs.
A free sudoku you can play or print.
A free maze generator, one solution guaranteed.
A word search made from your own word list.
Use cases A teacher prints a pack of sudoku or word searches that have each been independently checked for a valid solution. A puzzle-book seller generates a pack with a layout fitted to the page and a QA report. A parent prints mazes or logic grids for a rainy afternoon. Land on PuzzleForge all 4 screenshots, the full story What it is Python generators for sudoku, word search, mazes, logic grids and more, each with an independent solution validator, a print-fit layout engine and a storefront dashboard.
Highlights Every puzzle independently validated. Tests include deliberately corrupted puzzles. Three product definitions built and QA-reported. Built with Python ReportLab PyMuPDF Pillow pytest Static site Private repository · 2026
Working prototype · 2026 · last worked on Nov 4, 2026
RecallRadar
A searchable, self-updating site for official U.S. vehicle and consumer-product recalls.
One search across vehicle and product recalls.
Vehicle recall lookup, with the most-recalled brands since 2022.
The most urgent recalls, the do-not-drive advisories.
Comparing recall history across brands.
Use cases A car owner looks up a vehicle to see whether any official recall campaigns apply to it. A parent searches consumer-product recalls and adds items to a private watchlist that flags new recalls. An operator checks the freshness banner and ops dashboard to confirm the data is up to date. Land on RecallRadar all 4 screenshots, the full story What it is Combines tens of thousands of official recall records with search, vehicle lookup, watchlists and an operations dashboard, using no npm dependencies.
Highlights 30,000+ vehicle campaigns and 10,000+ consumer recalls ingested with provenance. Private watchlist that flags new recalls. Source-freshness banner and ops dashboard. 45 tests. Built with JavaScript Node.js Static site Data pipeline Private repository · 2026
Working prototype · 2026 · last worked on Nov 4, 2026
MacroKitchen
Free recipe-nutrition calculators and a food database built on USDA FoodData Central, plus spreadsheet products.
The front page: a food search and ten calculators.
A recipe built from two USDA foods, with an estimate flag where household measures are used.
Each calculator shows its math, with a worked example.
The food database, browsable by category.
Use cases A home cook builds a recipe and gets its nutrition totals from USDA FoodData Central values. Someone tracking macros looks up a food page to see its nutrient values and where they come from. A fitness coach uses one of the ten calculators to work out a number for a client plan. Land on MacroKitchen all 4 screenshots, the full story What it is A static site with calculators, a recipe builder and thousands of food pages, validated against independently computed reference cases.
Highlights Ten calculators and a recipe builder. About 7,800 generated food pages. 189 independent reference cases and 237 tests. Built with JavaScript Python Static site Node.js Spreadsheet generation Private repository · 2026
Working prototype · 2026 · last worked on Nov 4, 2026
Roblox Game Factory
A local HQ dashboard and AI-assisted pipeline that researches, builds, tests and tracks original Roblox games.
The factory's world in my Constellation command center.
The same world from above.
A closer look at the production ring and its games.
The HQ dashboard overview, from the repository's design notes, with demo data.
The games gallery, shown with demo data.
A game's detail page, shown with demo data.
The telemetry drawer, shown with demo data.
Use cases Alex queues a new original game idea and lets the headless pipeline build it with Claude Code. A developer checks that a build passes real in-engine tests run through Roblox Studio before release. Alex reviews the hash-chained audit log and dev-time tracking to see what the pipeline did and how long. Land on Roblox Game Factory all 7 screenshots, the full story What it is A dashboard with dev-time tracking and a headless job pipeline that builds games with Claude Code and tests them through Roblox Studio.
Highlights Produced the published game Wobble Bay. Studio test harness drives real in-engine tests. Hash-chained audit log and verified backups. Built with TypeScript Node.js React SQLite Rojo Luau Playwright Private repository · 2026
Working prototype · 2026
Glassline GT
A browser racing game alpha: a 100-car garage, 12 original circuits, career mode, tuning and photo mode.
The race menu over the live 3D scene, with some figures blurred.
Picking a circuit changes the 3D scene behind the menu.
Another circuit, another scene.
Use cases A racing fan browses the searchable garage and tunes a car for the next circuit. A player works through a career of championships across the four tours. A player uses photo mode to frame and capture a car on one of the original circuits. Land on Glassline GT all 3 screenshots, the full story What it is A standalone Three.js racing game with 12 original circuits across four tours, a searchable garage and per-car tuning saved locally.
A development alpha, not a finished game.
Highlights 12 original circuits in 4 tours. 100-car garage with filters and tuning. Photo mode and career/championships. Described in general terms only, on purpose. No screenshots, data or results appear on this site.
Built with TypeScript Three.js Vite Playwright Private repository · 2026
Working prototype · 2026
Firefly
A local autonomous procedural-art studio that creates a new 3D sculpture every fifteen minutes.
The current exhibit: a procedural sculpture made by the local service.
Another work from the archive.
A helix-form work from the archive.
Use cases A generative-art fan opens the viewer to see the newest procedural sculpture, made every fifteen minutes. A developer studies how deterministic per-slot generation yields the same sculpture whenever a slot is re-run. Someone browses the capped archive of past sculptures to compare them, with no accounts or outside services involved. Land on Firefly all 3 screenshots, the full story What it is A small service that deterministically generates and archives procedural curve sculptures, with a viewer and a read-only contract.
Highlights Deterministic per-slot generation. Capped atomic archive. No external APIs or accounts. Built with Node.js TypeScript Three.js Node test runner Private repository · 2026
Working prototype · 2026
Autonomy Labs
Three small autonomous generators: fictional star systems, synthetic islands and marble-track designs, each on its own clock.
Kepler: a new fictional star system every six hours, labelled synthetic.
Tideline: one synthetic island a day, with coastline statistics.
Kinetic: an hourly downhill marble-track design, an illustration rather than physics.
Kepler's world in my Constellation command center.
Use cases Someone curious about generative art opens the viewer to see the latest fictional star system, with its slot and date. A developer studies how a small service can make one new thing on a schedule and archive it without ever back-filling. A visitor looks through the island archive to compare coastline statistics from different days of synthetic terrain. Land on Autonomy Labs all 4 screenshots, the full story What it is Kepler, Tideline and Kinetic are tiny services that each make one new thing on a schedule, archive it and show it in a viewer. Everything they make is fictional and labelled synthetic.
Highlights Kepler: a new fictional planetary system every six hours. Tideline: an original island heightfield every day, with coastline statistics. Kinetic: a marble-track design every hour (an illustration, not a physics simulation). Each slot is generated once, never back-filled, and archived atomically. These live inside the Constellation repository as small services.
Built with Node.js TypeScript Three.js Node test runner Private repository · 2026
Working prototype · 2026
Content Studios
Three automated studios that draft original SVG icon packs, tabletop encounter maps and landing-page templates.
The icon studio's page, with names, previews and figures blurred.
The map studio's page, blurred the same way.
The template studio's page, blurred the same way.
Further down: a download for the owner and a note that nothing is for sale yet.
Use cases A designer reviews a draft icon pack, with usage notes, before deciding whether it is worth polishing. A game master looks over a draft encounter map and its guide before anything is considered for release. Alex checks a draft landing-page template in the unlisted queue, since nothing is published until he approves it. Land on Content Studios all 4 screenshots, the full story What it is Vektoryard, Questtile Press and Foldlaunch Kits generate and package original digital bundles on a schedule without paid services. Everything stays an unlisted draft until it is reviewed.
Drafts only. Nothing is listed or sold.
Highlights Vektoryard: coherent 24-icon SVG packs with usage notes. Questtile Press: original encounter maps with game-master guides. Foldlaunch Kits: self-contained responsive landing-page templates. Nothing is listed or sold; the names are working titles. Described in general terms only, on purpose. No screenshots, data or results appear on this site.
Built with Node.js TypeScript SVG HTML/CSS ZIP packaging Private repository · 2026
In progress · 2026 · last worked on Nov 6, 2026
alextrosen.dev
This portfolio: click a hand-built rocket, blast off, and explore a small universe of the things I've built.
The front door of the site is a rocket you click to launch.
The home page after the launch, with the ringed planet and the links into the site.
The About page, with sticky notes on how the site was made and what is on the workbench.
The Now page, a short list of what I am working on.
Use cases A recruiter with two minutes uses the fast path to see what Alex has built and read a plain case study. A visitor who enjoys the rocket launch flies through the galaxy and picks a project to read about. A hiring manager shares one project's case study link with a teammate who has not seen the rest of the site. Land on alextrosen.dev all 4 screenshots, the full story What it is A static Astro site with a Three.js launch scene, plain shareable case studies and a recruiter-friendly fast path.
You're looking at it. Not published yet.
Highlights Case study per project. Privacy scan on every production build. Desktop and phone test suites. Built with Astro TypeScript Three.js Playwright Private repository · 2026
In progress · 2026 · last worked on Nov 4, 2026
AIRWAVE
A hand-tracked, ten-finger beat-making workstation that runs in the browser.
The studio with no camera on: ten finger pads, each mapped to a drum or synth sound.
Demo hands feeds synthetic landmarks through the same gesture pipeline, so you can watch it work without a camera.
Projects and samples: everything saves in the browser and exports as a project file.
Use cases A hobbyist with a webcam and no drum machine taps out a beat in the air using all ten fingers. A musician without a MIDI controller plays the same synthesized kits from the keyboard or mouse and records the take. Alex tests the gesture pipeline with demo hands, so he can check new sounds without a camera. Land on AIRWAVE all 3 screenshots, the full story What it is Webcam hand tracking and keyboard/mouse input feed one gesture pipeline driving synthesized drum kits, a sequencer and recorder.
Highlights Original synthesized sounds (no licensed samples). Shared pipeline for keyboard, mouse, demo hands and camera. Unit tests for audio, gestures, identity, sequencer and storage. Built with TypeScript React Vite MediaPipe Web Audio Vitest Private repository · 2026
In progress · 2026
Pocket Planets
A Roblox game in development at the Roblox Game Factory.
The system view: six tiny planets around your star.
Each planet explains why it is changing, reason by reason.
Visiting a world to meet its inhabitants.
Its world in my Constellation command center.
Use cases Alex checks the in-engine test results to see how far the game has come before the pipeline resumes. A reader following the factory sees how a second original Roblox game moves from idea to tested build. Alex uses the version 0.3.0 repo as a baseline when the build pipeline is switched back on. Land on Pocket Planets all 4 screenshots, the full story What it is A second original game being built by the factory pipeline; not published.
Not published. Its in-engine tests pass, but the build pipeline is paused.
Highlights Version 0.3.0 in the repo. Large automated in-engine test suite passing in Studio on 2026-10-26. Built with Luau Roblox Studio Rojo Automated Studio tests Private repository · 2026
In progress · 2026
Ramen Runners
A Roblox game in development at the Roblox Game Factory.
The ramen street, built by the game's own world builder.
Service: rush orders queue up along the top.
The kitchen planner routes each dish through the stations.
Its world in my Constellation command center.
Use cases Alex checks the in-engine test results to see whether version 0.1.0 is ready for further work. A reader following the factory sees how an original Roblox game is built by a coding-agent pipeline. Alex resumes the game from the repo when the paused build pipeline is switched back on. Land on Ramen Runners all 4 screenshots, the full story What it is An original game being built by the factory pipeline; not published.
Not published. Its in-engine tests pass, but the build pipeline is paused.
Highlights Version 0.1.0 in the repo. In-engine test suite passing in Studio on 2026-10-26. Built with Luau Roblox Studio Rojo Automated Studio tests Private repository · 2026
In progress · 2026 · last worked on Oct 26, 2026
The Great Detour
A planned original visual exploration game, The Explorer's Realm.
Use cases Alex uses the scoped first demo as a plan for a valley, canyon and ridge route with a character controller. A follower of the portfolio sees what a planned original exploration game would involve. Alex refers to the reserved repository when deciding whether to start building. Land on The Great Detour the full story What it is A future game project. The first demo is planned (valley, canyon and ridge route with a character controller); nothing is built yet.
Planned. Nothing is built yet.
Highlights Repository reserved. First demo scoped. Built with Private repository · 2026
In progress · 2026 · last worked on Nov 3, 2026
Big Data Bowl 2027 entry
A research pipeline testing whether combine sensor tracking reveals movement abilities that stopwatch times miss.
The scouting desk: 337 draft prospects from the competition's data.
A workout card, with measured values and honest not-tested markers.
The methods page says where every number comes from.
The Motion Lab replays recorded tracking frame by frame.
Athlete Twins finds players who tested like this one.
Use cases A competition entrant runs the full pipeline to test whether combine tracking data shows what stopwatch times miss. A reviewer checks the claims register and locked holdout to see exactly what the results are allowed to say. Alex uses the private scouting app built on the pipeline to explore the data locally. Land on Big Data Bowl 2027 entry all 5 screenshots, the full story What it is A reproducible analysis pipeline and write-up for a public analytics competition, with a locked holdout and a claims register that fixes the wording of results.
Highlights Pre-registered research protocol and locked holdout. Private local scouting app built on the pipeline. Submission deadline 2027-01-06. Built with Python DuckDB Polars pandas statsmodels scikit-learn Private repository · 2026
In progress · 2026 · last worked on Nov 2, 2026
Safety tooling for cleaning up and customizing a household streaming device.
The toolkit's own usage text: every change is per-user, logged and reversible.
The removal plan, read from a local list without contacting any device; counts blurred.
Use cases Before changing a household streaming device, the owner takes a snapshot so every change can be undone. Someone staging customizations runs health checks after each step to confirm the device still works. A cautious tinkerer uses the full-restore tooling to return a device to a known good state. Land on Device Tune-Up Toolkit the full story What it is A small toolkit built to make changes to a device reversible and well-checked. Work is on hold.
On hold. Nothing has been changed on the device.
Highlights Snapshot and full-restore tooling. Staged changes with health checks. Described in general terms only, on purpose. The screenshots are the tool's own usage text and removal plan, made without contacting the TV; every count is blurred and no device, address or account appears.
Built with Python Shell Android Debug Bridge Private repository · 2026
In progress · 2026
Forge Website Builders
A small-business web design service concept.
The marketing page, with the logo and prices blurred.
Package cards and recent work, with prices and names blurred.
The portfolio grid, with every business name blurred.
Use cases A small-business owner compares pricing tiers to see what a simple website would cost. Alex shows sample site designs and a brand kit to a prospective customer as part of a pitch. A designer reuses the templates as a starting point for a small-business page. Land on Forge Website Builders all 3 screenshots, the full story What it is Business materials, pricing, templates and demo pages for a web design service. Not an active software product.
A business concept with materials, not a running product.
Highlights Brand kit and templates. Pricing tiers and sample site designs. Described in general terms only, on purpose. No screenshots, data or results appear on this site.
Built with HTML/CSS Spreadsheets Documents Private repository · 2026
In progress · 2026
VoxelMint
A planned studio for building and releasing Minecraft Fabric mods.
Its world in my Constellation command center; the app is only planned.
The same planned world from above.
A closer look; the stations show no data because nothing is built yet.
Use cases Alex uses the documented design as the starting point if he decides to build a mod studio. A reader sees what an idea-to-release pipeline for Minecraft Fabric mods would cover, though nothing is built yet. Alex keeps the reserved Constellation world for this future project in view. Land on VoxelMint all 3 screenshots, the full story What it is A future project: an idea-to-release pipeline for game mods. Nothing is built yet.
Planned. Nothing is built yet.
Highlights Constellation world reserved. Design documented. Built with Java (planned) Fabric (planned) Private repository · 2026
Working prototype · 2026 · last worked on Nov 4, 2026
Blueprint Zero
A rough sketch and a sentence become an editable multi-part CAD assembly with real moving joints.
The start page: a sketch canvas, a sentence box and a demo button.
After the demo, the parser lists what it understood and offers three mechanism templates.
Strokes can be tagged as base, arm, rest or hinge before anything is built.
The built phone stand: a multi-part 3D concept with real joints and editable dimensions.
Checks, parts and the STEP and STL export, with a concept-design warning on every page.
Use cases A maker sketches a phone stand with two hinges and gets jointed parts to adjust and export. A tinkerer types 'make the base 2 cm wider' and reviews the change as a diff before applying it. Someone preparing a 3D print downloads STEP and STL files that were reopened and measured first. Land on Blueprint Zero all 5 screenshots, the full story What it is A local sketch-to-invention workbench: rule-based intent parsing proposes a mechanism, CadQuery builds the parts and joints, plain-English revisions apply as a diff, and STEP/STL exports are reopened and measured before they're called verified.
Concept design only: it computes geometry and rigid-body motion, not strength, loads or print time.
Highlights Three parametric templates with revolute and prismatic joints, clamped to their limits. Feasibility checks on exact solids: collisions, thin walls, joint limits, build volume. Every number carries a unit and a source badge: measured, user-entered, estimated or verified. Built with Python CadQuery 2.8 FastAPI React Three.js TypeScript Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 4, 2026
MAKESHIFT
A broken knob, spacer or clip plus a few caliper measurements become a printable replacement part.
The start page: describe the break and give a few measurements.
A replacement knob in 3D with a parameters panel and a fit contract.
A validation report checks fit and printability, and marks sizes as measured or estimated.
Use cases A homeowner whose lamp knob snapped off measures the 6 mm D-shaft and prints a replacement. Someone missing a spacer enters shaft and length measurements and gets a part with honest clearance. A 3D-printing hobbyist prints the three-bore calibration coupon to tune fit before the real part. Land on MAKESHIFT all 3 screenshots, the full story What it is A local replacement-parts studio: a rule parser reads the description, CadQuery builds a real parametric solid, a worst-case fit check names any interference, and exports include a calibration coupon and a fit-test checklist.
Concept design: needs physical verification before use. Refuses mains, gas, vehicle, medical and other safety-critical parts.
Highlights Knob, spacer and spring-clip families, each with typed parameters and limits. Fit verdict at the worst case, with the overlapping solid highlighted and a one-click repair. Rejects a part with no measured anchor and hands back a caliper checklist instead. Built with Python CadQuery 2.8 FastAPI React Three.js TypeScript Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 4, 2026
SKIN
An object's real dimensions plus a style prompt become a printable sleeve with locked fit and openings.
The start page: enter measured sizes or try the demo.
The demo shell in 3D, with the cavity and a connector opening.
Every dimension is labelled measured, user-entered or estimated.
Use cases Someone with a small gadget enters its size and gets a printable sleeve that keeps the ports open. A maker compares industrial, matte and brutalist exteriors on the same fit before printing. A hobbyist prints the corner coupon to check clearance before committing to the full shell. Land on SKIN all 3 screenshots, the full story What it is Builds a rectangular shell around an object you measure. The cavity comes only from the size and clearance, keep-out openings for ports and buttons are cut last so no style can cover them, and STL, STEP and SVG exports are verified.
First slice: rectangular sleeves only. Cylinders, clamshells and snap features are planned.
Highlights Three exterior styles shown side by side on the identical fit. Export is blocked when a wall measures thinner than the print profile allows. A 1:1 SVG drawing with outer and cavity dimensions and the keep-out outlines. Built with Python CadQuery 2.8 FastAPI React Three.js TypeScript Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 4, 2026
Oracle
Will it rain tomorrow? A forecasting tournament scored honestly on ten years of real station data.
A tournament of three rain forecasters scored on real station data.
Calibration and score breakdowns under the leaderboard.
The forecasts that were most wrong, with what each model had seen.
Live forecasts written to an append-only ledger before they close.
Use cases A forecasting learner compares models on 88,934 scored city-days and sees where each one fails. Someone checking a claim of accuracy verifies the ledger's hash chain to confirm nothing was edited. A data student filters results by city, season and year to see how calibration changes. Land on Oracle all 4 screenshots, the full story What it is Asks one exact question in 35 cities and answers it three ways: a walk-forward tournament on recorded NOAA data, an append-only hash-chained ledger of forecasts made before each day closes, and scores checked against an independent implementation.
Weather only: no trading, no markets, no investment advice.
Highlights Each test year is forecast only from the years before it. Brier score, log loss, calibration and bootstrap confidence intervals. A failures page with the most confident misses and what each model could see. Built with TypeScript Node.js Python NOAA GHCN-Daily Vitest Playwright Private repository · 2026
Working prototype · 2026 · last worked on Nov 4, 2026
Sentinel
Break a toy service on purpose and watch Sentinel find the cause, without ever repairing it for real.
Three toy services, all healthy, with an earlier incident shown as resolved.
A drill stops a toy service and Sentinel finds it on its own, then dry-runs a fix.
A bad config is detected and a revert playbook is dry-run, awaiting approval.
Use cases A developer learning incident response takes Inventory down and follows the evidence to the root cause. A reliability engineer reviews a playbook's preconditions and blast radius in a safe dry run. A team rehearses writing a postmortem from a real, recorded toy incident. Land on Sentinel all 3 screenshots, the full story What it is Observability and dry-run self-healing for three synthetic services it starts itself: a chaos button injects a fault, Sentinel links logs, probes and process events into one incident, names the likely cause with evidence, and plans an allowlisted fix it refuses to execute.
Real execution is disabled by design; Sentinel never touches anything it did not start.
Highlights Four injectable faults on real child processes, bound to 127.0.0.1. Hypotheses from a fault taxonomy with evidence chips and a confidence score. A hash-chained audit log and a postmortem export for every incident. Built with TypeScript Node.js React Vitest Playwright Private repository · 2026