AI product engineering
Agents that use tools, evals that gate releases, and costs kept on a leash. Built for paying users, not demos.
- level
- strong
- years
- 2
- evidence
- work/pixro, work/real-estate-gpt
What I can do
- Build agents that call tools to get real work done: Pixro’s WhatsApp and Instagram shop assistant takes a customer’s message in any of several languages, decides what to make, and generates images and videos through tool calls.
- Test AI behaviour properly. I wrote an eval harness that runs the real agent against a real model, grades its decisions, and fails the build below a pass-rate threshold or on any guard-rail miss. It never touches the database, WhatsApp or credits, so it’s safe to run anywhere.
- Keep AI costs sane: a stronger model plans and a cheaper one writes, long system prompts are cached, tokens are capped, and a cheap default model is used where it’s good enough.
- Give agents memory and judgement: a rolling conversation summary plus recent turns, a durable brief, optional vector recall, and skills loaded only when they’re needed.
- Build retrieval and memory at scale: for Real Estate GPT, a property assistant with over a million users, I built on LangChain, LangGraph, AWS Bedrock and Qdrant, with streaming chat, persona-based memory and long-term retrieval.
- Design human-in-the-loop flows: the blog agent streams its reasoning and proposes edits as diff cards you approve or reject, then scores the draft on a 10-point rubric (rules plus an LLM judge).
How I learned it
By building Pixro and consulting on Real Estate GPT. Every one of these started as a problem with a real customer or a real bill.
Where it shows
Pixro: the multi-channel shop assistant, the blog-writing agent and the content scorecard are all modules I wrote. Real Estate GPT: RAG and memory at scale.