I build AI platforms by day and AI side quests by night.
My rule: the best way to understand something is to build it. So when I wanted journaling to feel effortless and visual, I built an app. When I wanted a second opinion on stocks, I built a committee of AI analysts. When I wanted learning algorithms to feel like play, I built a game for it.
A committee of AI analysts that lives in my WhatsApp.
I message my agent analyze AAPL, and about 10 minutes later a full research report arrives.
- 🧮 Formulas do the math, the LLM only interprets it. Valuation models, quality scores and technical signals are deterministic and unit-tested.
- 🪑 Seven analyst seats (fundamental, valuation, technical, sentiment, a Buffett lens, a Burry lens, macro) vote, and a citation check flags any number that isn't in the data.
- 😈 The trader has to argue against the majority first, then three risk managers (aggressive, conservative, neutral) push back on its plan. An optional bull/bear debate can be switched on per run.
- ⚖️ Two judges plus an "ablation judge" that decides without the fair-value estimate. If they disagree, confidence is capped.
- 📉 It scores its own past calls against real prices and against plain monthly investing. My own backtest found the macro tilt doesn't predict returns, so the report says so.
Python · OpenClaw · SEC / FRED / Yahoo data · 350+ unit tests
Research only, not financial advice.
🎬 Scène
Turn your day into a movie, one line at a time.
Write one line about your day and get back a cinematic scene, a quote and a song that echo it. Over time your scenes collect into a film roll of your life. 🎞️
- Built and shipped solo on iOS
- Firebase + App Check on the backend
- An agentic pipeline creates the weekly social posts and ads, so marketing is automated too 🤖
Swift · Firebase · Generative AI
What if a whole dev team were agents?
Nine agents (PM, Planner, Coder, Reviewer, QA, DevOps, Researcher, Data Analyst, and a Prompt Optimizer that tunes the others' prompts) work together behind an LLM router.
- Orchestration modes: router-decides, sequential, parallel, supervisor and auto
- Human-in-the-loop checkpoints, a RAG pipeline and a visual workflow builder
- DeepEval regression tests on golden datasets, plus prompt-drift detection
- Full OpenTelemetry tracing with a live trace inspector 🔍
LangGraph · MCP · FastAPI · Next.js
Practising data structures and algorithms, but make it a game.
- Algorithm visualiser with real-time animations synced to the Python code
- Challenges across sorting, graphs, trees, DP, greedy and bitwise
- XP, achievements and progress tracking (and dark mode, obviously 🌙)
Next.js · TypeScript
I like fixing the tools I actually use.
- OpenClaw: four fixes landed. My financial agent runs on it, so I'm also a user.
- Discord messages silently dropped during gateway reconnects (#100896)
- Google Chat streaming settings lost when
doctormigrates named accounts (#106018) - A websocket cache race on stale reused releases (#105634)
- Control UI run status getting stuck on stale runs, landed by a maintainer as a co-authored rework (#100527)
- DeepEval: fixed missing token counts and costs for GPT-5.x in LLM tracing spans, with unit tests (#2601). I use DeepEval for agent evaluation at work, which is how I found the bug 🐛
I'm an AI platform engineer working on enterprise GenAI infrastructure:
- AI Gateway: one API to 80+ LLM, embedding and multimodal models, with routing, caching, rate limiting and guardrails
- Agent Hub: a low-code platform for building multi-agent teams, with built-in evaluation
- Agent sandbox: isolated, ~100ms-to-start environments where agents can run untrusted code safely
AI: LangGraph · LangChain · MCP · DeepEval · RAG · PyTorch fine-tuning Languages: Python · TypeScript · Go · Swift · SQL · Lua Build: React · Next.js · FastAPI · PostgreSQL · Redis · Firebase Infra: Kubernetes · Docker · Terraform · AWS · GCP · Azure
AI gateways · multi-agent design · LLM evaluation · why my AI stock committee argues with itself

