Agent systems, trading research and LLM pipelines. Most of them exist to answer one question honestly — does this actually work? — even when the answer is no. Source code available on request.
A deep-analysis engine that summons 38 philosophers, East and West, to read the world's events back to us in their own voice. Citation-anchored and bilingual. It is the engine behind the essays on this site.
Read the essays →Pick a topic and a cast of philosopher personas, then watch them respond to it and to each other over multiple turns.
A 26-week mechanistic interpretability course, run inside Claude Code: lessons and coding exercises on disk, tests, and an AI grader that checks the work.
An AI investment team: nine agents analyse, debate and produce educational investment reports.
Four specialist analysts — technical, on-chain, sentiment, macro — each write a report on a coin. A bull/bear debate and a trader step turn those reports into a decision.
Autonomous, AI-driven strategy research for crypto perpetual futures on Hyperliquid, in the spirit of karpathy/autoresearch.
Reality-check scripts that test whether a strategy survives realistic costs, plus a risk-gated paper-trading bot. The honest conclusion: nothing tested was reliably profitable after costs.
Every 30 minutes it pulls the Hyperliquid all-time-PnL leaderboard, filters out market makers and vaults, reads what the remaining traders hold, and alerts on Telegram when they converge.
A phased build of a quant stack on Futu OpenD — connection proof first, then data, backtests and paper execution.
Copying a good trader is mostly a plumbing problem: find wallets worth following, notice their trades quickly, and decide whether each trade is still worth copying.
News-driven edges in prediction markets are easy to imagine and hard to prove, so the proof is the product: a narrow watchlist, paper results, and a local dashboard.
Two automated monitors: a twice-daily scan for wallet and market alerts, and a copy-trading intelligence feed with Telegram alerts, both running free on GitHub Actions.
An auditable scenario simulator for departmental P&L: Monte Carlo runs, a Streamlit dashboard, and every input traceable to its source.
A daily email digest of trending Python repositories, curated by Claude into actionable buckets.
An experimental worldview and workflow product prototype, taken from PRD through wireframes to a build-ticket breakdown.
This site: daily essays and weekend roundtables, built in React and deployed on Cloudflare Workers.
Visit →A crash-game simulator with a full game loop — waiting, countdown, flight, crash. Simulation only; no real money.
The design bible for a fantasy courtroom-mystery visual novel: world, characters and case scripts.
A web-based search assistant for bug bounty hunters and security researchers.