I don't use AI tools — I build them. AI is the medium I work in, not a tool I borrow.
- AI engineering — design and ship MCP servers, multi-agent systems, and agent fleets. Author of two patterns I keep reusing: a four-file agent-identity stack (soul · brain · playbook · memory) and sub-agent isolation with thin orchestration (no logic in the orchestrator).
- Agent design — composable, named skills over monolithic prompts;
_base/templates that make a new domain agent a week's work, not a month's. - Applied AI research — treat failure modes as research problems, not prompts to tweak. Diagnose the ceiling before touching the model.
- Creative & product — name the thing, give it a soul. Identity-aware copy, distinct-aesthetic interfaces, narratives people feel.
- Computer vision — image scoring, feature extraction, automated ranking.
| Project | What it is |
|---|---|
| wealthsimple-mcp | MCP server wiring agents into the Wealthsimple Help Center — 7 typed tools, Zod-validated boundaries, opt-in observability. No scraping, no auth. TypeScript. |
| linkedin-mcp | Authenticated LinkedIn MCP server — Voyager API + data-export posts. TypeScript + Python FastMCP. |
| notion-template-generator | MCP server for Notion automation — 39 tools across 8 categories: wiki management, content generation, AI-powered reorganization. Python. |
| future-of-work | AI-built research knowledge base on the future of work — 6 domains, 400+ articles. Built with RivalSearchMCP + Cursor, deployed on GitHub Pages. |
| metaposters | An independent, AI-native poster studio. metaposters.ai |
| poster-analyzer-and-picker | Computer-vision tool that scores and ranks poster designs on color, texture & composition. Python. |
| Languages | Python · TypeScript · JavaScript · Rust · SQL · HTML / CSS |
| AI & agents | MCP (protocol + server design) · Claude Code (skills, sub-agents, slash commands) · Cursor · FastMCP · multi-agent orchestration · RAG · prompt engineering · tool use |
| Data & ML | pandas · scikit-learn · Jupyter · feature engineering · hypothesis testing & experimental design · computer vision |
| Web & deploy | Node.js · FastAPI · Jekyll · GitHub Pages · GitHub Actions |
| APIs & integrations | Zod · Notion API · Slack API · Datadog (LLM Observability) · GraphQL · BigQuery |
Learning: NVIDIA AI Infrastructure & Operations · BrainStation Data Science (certified)
Build → codify → cite → diagnose → measure.
Every workflow I run twice becomes a reusable skill. Every claim ships with a source. When an AI system stalls, I treat it as a research problem — not a prompt to tweak.