AP-2002

A Rust runtime that hands an agent the same interfaces a person uses: screen, mouse and keyboard, a browser over CDP, a terminal, the filesystem. Every capability goes through a policy engine that allows it, denies it, or holds it for approval, and each action lands in a hash-chained append-only log, so you can read back afterward what the agent actually did. Tasks, memories, decisions, and permissions live in SQLite instead of a conversation transcript. Twelve crates sit behind one trust boundary. A Tauri 2 desktop app and a CLI ride on top of the same runtime and carry no agent logic themselves. Model providers swap out: Anthropic, OpenAI-compatible, and a mock for tests.

A sportsbook-shaped odds platform running entirely on hardware I control. Odds come from a licensed data provider, get normalized and priced, then stream to the browser over a WebSocket. Nothing polls anywhere, and there is no fixture data, so an empty board means an empty pipeline. Six Go binaries out of one module move events through Kafka, with Postgres and TimescaleDB holding the line history and Redis holding the current snapshot and socket state. The odds library is pure functions under property-based tests: American, decimal, and fractional conversion, four devigging methods, no-vig fair value, expected value, Kelly sizing, closing line value, and parlay pricing with same-game correlation. Balances are play money. No real funds move, no identity checks, no custody.

X's core ranking pipeline, rebuilt as a tunable FastAPI service. Candidates score on five weighted signals, popularity, recency, diversity, graph affinity, and semantic similarity, and each weight is a live parameter in the Next.js interface, so moving one reorders the feed in front of you. A simulation layer on Gemini and OpenAI writes synthetic personas and posts to fill the system and push on it, while MiniLM sentence embeddings match candidates against a user's recent likes by cosine similarity.

A Model Context Protocol server that fetches a URL, strips the ads, navigation, and boilerplate wrapped around the content, and hands back plain text. Scraping moves inside the tool ecosystem the model already speaks, so nothing has to be pasted in by hand.

An interactive dashboard that holds DAO governance up against the corporate kind. It reads voter turnout, concentration of voting power, and how long decisions take across Uniswap, MakerDAO, and Aave, then sets those numbers beside Meta and Apple. Built from participation charts, power-distribution visualizers, and guided walkthroughs of on-chain versus off-chain voting. The stubborn part was normalizing data across five DAOs that each count a vote differently.

Connect Four in C on an STM32F429I Discovery board. The ILI9341 LCD draws the grid and the STMPE811 controller reads the touch, so you slide a finger across the panel to pick a column and drop the piece. Win detection sweeps horizontal, vertical, and both diagonals out from the piece just played instead of rescanning the whole grid. The opening menu offers two players or one against the board, where the hardware RNG picks a legal column with no evaluation of the position.

Forty RISC-V instructions running on a five-stage pipeline: fetch, decode, execute, memory, write-back. ME/WB-to-ID operand forwarding clears the data-hazard stalls, interlocks catch the load-use case forwarding cannot, and branches resolve in memory behind a two-cycle flush, with branch prediction covering the common path. Tuning the stages bought close to 8% in clock-cycle efficiency. Written in Codal at CU Boulder.

A social platform built around playlists rather than posts: people assemble them, share them, and dig through other people's. Song data comes from music APIs and hangs off user profiles, and the usual social layer sits on top, following, likes, and comments. Still in progress, with no public repository yet.