ijtihed.

Description

An LLM-first fighting game. Every move grid, hitbox, and visual effect is data the LLM can author at runtime.

why I built it

always wanted a fighting game where you could play as literally anything you can describe. felt like that was finally possible with llms.

Description

An agentic-first game engine. Open source, in progress.

why I built it

everyone in the ai lab was building web games, but nobody was utilizing gpu / unity cuz it was built for humans not agents. this changes that. an ai native game engine.

Description

~2.2k lines of Rust, no game engine, single GPU shader. Chunked dirty-rect scheduler (the Noita optimization), 10 cell kinds with density-based displacement and reactions, history ring buffer with timeline forking, hand-written WGSL bloom + Reinhard tonemap, 18 scenarios. ~180M cells/sec saturated. 19 tests + CI on Linux/macOS/Windows.

why I built it

wanted a project that exercises systems programming, low-level graphics, randomized cellular automata, and a non-trivial UI all at once, but small enough that one person can hold the entire codebase in their head.

Description

Lagrangian mechanics for three point masses on rigid rods, 3x3 mass matrix solved by Cramer's rule with six shared 2x2 subdeterminants, classical RK4 at dt=1e-4. State and scratch held in 80-bit extended precision via x87; six fsincos per derivative call. Terminal renderer at 30 fps with ANSI cursor escapes. Browser port reuses the same algorithm with a double/triple toggle and per-phase asm highlighting. Beats SciPy DOP853 in energy conservation at long times by ~10x by keeping state in 80-bit.

why I built it

filtered language:Assembly + triple pendulum on github. zero. tried double, n-pendulum, multi-pendulum. zero. sourceforge, gitlab, codeberg, rosetta code, pouët, demozoo. zero. the math is in every classical-mechanics textbook. nobody had bothered to write it in asm. so this exists.

Description

An IDE for the architecture of your codebase. Describe a change in plain English; schema turns it into AST transforms, runs them through your test suite, and animates the dependency graph as it mutates.

why I built it

ides show you files, not architecture. you can stare at thirty open tabs and still not see the shape of the system. wanted a view that thinks at the level of clusters and routes.

Description

GraphRAG over Finnish tax law (Finlex, KHO, Vero; 402k chunks, 1.97M graph nodes). Citation-backed answers with a live reasoning graph; vector baseline vs. graph-routed + cross-encoder rerank ablation.

why I built it

taxxa.ai challenge at the aaltoes prompt finance hackathon: agentic GraphRAG for accounting regulatory research. team build over the weekend at villa upinniemi. first place — $3,000+.

Description

TBD

why I built it

none of the big platforms had instant anki-style puzzle generation for learning openings and mistakes, so had to make one myself.

Description

Internal cohort playtesting tool for Supercell: built to replace ad-hoc sharing; now the AI Lab's pipeline for co-testing.

why I built it

a quick batch meeting showed me that everyone wanted everyone to playtest their games but nobody wanted to just drop a random link in slack; built this to support everyone in the batch and make it easy to get feedback and iterate quickly on it.

Description

A historical simulation project. Very early stages, more details coming.

why I built it

work in progress.

Description

23 browser-based simulations (cellular automata, particle systems, and artificial life), each built from the source papers. No build step, no dependencies; Canvas, typed arrays, and requestAnimationFrame.

why I built it

was talking with my friend about game of life, and i thought about how there must be millions of new papers that were never visualized that could be super cool. so i built this in an afternoon.

Description

npm package that cross-references every claim in your README against actual files, CLI flags, env vars, badges, and function signatures. Catches drift before users do. Run as a CLI (npx driftmd), GitHub Action, or AI prompt. Zero config.

why I built it

believe it or not, nobody has done this before. weird. well since it didn't exist and i needed it, i made it.

Description

A VS Code extension that replaces the standard breakpoint view with a canvas where state is directly editable. Fork into up to four parallel timelines, run each forward independently, and commit one to extract a code fix that would have produced that state without manual intervention.

why I built it

debugging is a search problem but every debugger only lets you sample one point at a time. wanted one where the search itself is the editing surface.