Before I let coding agents anywhere near a production system, I wanted to actually learn to drive them. So I pointed them at my own life first — the decisions I kept overthinking, the spending I kept ignoring, the startup ideas I kept half-validating — and vibe coded a working product for each one.
That's this repo. Every project is a small, self-contained app — a FastAPI backend, LLM agents doing real work (live web research, adversarial analysis, transaction classification), and a UI on top — built with the same discipline I'd want on a real team: CLAUDE.md contracts, custom subagents, hooks, and test suites that keep the AI honest.
Each project's README has a demo video recorded from a real run — no mockups.
Roast Wallet — Your spending, on trial. Upload a bank statement and an AI roasts your month, grades it 0–100, and hands you a no-excuses plan for the next one. Or hook it up to your real bank and let the Butler watch your budgets and email you before a category blows.
FastAPI · GPT-4o-mini · Nordigen · Resend
Decision Foundry (Prism) — Pressure-test a startup idea before you build it. Your concept gets forked into 4 strategic universes, each torn apart by 6 web-researching agents, then graded against YC, a16z, Sequoia, and Techstars.
FastAPI · GPT-4o-mini · DuckDuckGo · React · SSE
Devils Advocate — That decision you've been avoiding — fed to five AI agents who research it, attack it from every angle, and refuse to leave without a verdict: GO, NO-GO, or NOT YET.
FastAPI · GPT-4o-mini · DuckDuckGo · vanilla JS
Debate Club — Two AIs go to war over any topic — AXIOM for, PHANTOM against — citing live web sources, getting flagged for fallacies mid-sentence, every claim mapped onto a D3 force graph. A third AI calls the winner.
FastAPI · GPT-4o-mini · Tavily · D3.js · SSE
The agentic tooling isn't hidden away — it's checked in alongside the code, because learning to wield it well was the whole point:
CLAUDE.mdcontracts — every project pins its architecture, module boundaries, and hard rules for the AI (where OpenAI calls are allowed to live, function-length limits, what never to write)- Custom subagents — Debate Club ships a
slop-detectoragent with persistent memory that reviews changes for AI-slop patterns and learns the codebase's conventions over time - Hooks — Decision Foundry auto-runs its test suite and ruff after every AI edit, blocks any shell command that would expose secrets, and writes a session summary when work stops
- Slash commands — repeatable workflows (
/lean,/tune-agent,/review-output,/add-search-source) live in the repo, not in someone's head - Real test suites — mocked unit tests for speed, plus live agentic evals that verify the agents actually search the web and ground their answers (Devils Advocate alone runs 175 checks)
They all follow the same pattern:
cd <project>
pip install -r requirements.txt # Decision_Foundry lists deps in its README
cp .env.example .env # add OPENAI_API_KEY (Debate Club also needs TAVILY_API_KEY)
uvicorn <app module> --reload # exact command in each project's READMEOpen http://localhost:8000. Each project README has the full setup, API reference, and test instructions.