I'm a Technical Product Manager with a background in applied physics and engineering, interested in how AI is redefining product management, not just as a tool but as a shift in how products are built, decisions are made, and roles are evolving.
I focus on developing technical depth across data systems, machine learning, and infrastructure, with the belief that strong PMs need to understand the systems they are shaping. I am interested in how AI can increase productivity and change workflows, where it meaningfully improves decision-making, and where it falls short.
I analyze emerging trends and products in the AI space and reflect on the realities of product work, from ambiguity and trade-offs to the gap between frameworks and execution.
My goal is to operate as a builder-first product manager, connecting deep technical systems with real-world impact.
| Project | Description | Tech |
|---|---|---|
| Customer Interview Synthesizer (n8n) | Multi-agent n8n workflow that turns raw customer-interview transcripts into prioritized product insights. Four worker agents (themes / JTBDs / contradictions / quotes) run in parallel → synthesizer → LLM-as-judge scores against a hand-curated reference → either auto-save or pause for Slack human-in-the-loop approval. The trust scaffolding (judge + HITL gate + planned drift monitor) is the artifact's main differentiator. | n8n Cloud, OpenAI GPT-4o-mini, Google Sheets, Slack |
| Nonna's Cookbook RAG | Agentic recipe assistant inside my iOS app — three-path routing (personal knowledge base, LLM general knowledge, live web search) gated by an LLM-as-judge relevance check. Built-in Observer mode visualizes each pipeline step in real time. Fully instrumented with LangWatch, scored against a hand-labeled golden dataset by three evaluators (routing correctness, gate correctness, LLM-as-judge answer quality) | Flutter, Firebase Cloud Functions, Firestore vector search, OpenAI (GPT-4o + text-embedding-3-small), Tavily, LangWatch |
| Model Benchmark | Blind A/B/C test of frontier LLMs on real PM tasks (synthesize feedback, write briefs, prioritize, analyze competitors, notes-to-decision) — reveals which model wins which workflow, not which wins generic benchmarks | Next.js 16, TypeScript, Tailwind, Anthropic + OpenAI + Google + OpenRouter SDKs, SQLite |
| Product Discovery Agent | AI agent that researches a problem space like a senior PM — with configurable scoring, multi-source input, and a two-layer eval framework (execution + outcome quality) | Python, LangChain, GPT-4o-mini, Tavily, Pydantic, LangSmith |
| Token Battery | macOS menubar app that monitors Claude Code token usage in real time — like a battery indicator for your AI budget | Python, rumps, Claude API |
| LinkedIn Post Finder | Engagement dashboard to track and interact with influential PM/tech voices daily | Python |
| iOS App | iOS application (in development) | Swift |
| TensorFlow Articles | Supporting material for my TensorFlow articles | Python, TensorFlow |
| Coursera Capstone | Data science capstone project | Python |