Abdullah A.

I build LLM systems you can trust with the answers they give. theproducts.ai is mine end to end. It scrapes YouTube review transcripts and Reddit threads, runs claim extraction and sentiment through an LLM, then publishes buying guides with citations. Two rules hold it together: a claim only ships if it traces back to a source snippet, and the model summarizes numbers, it never produces them. Python and FastAPI behind it, Supabase for storage, and Gemini, Claude or OpenAI depending on what a given step actually needs. Most RAG demos work fine. What breaks in production is retrieval that returns something plausible instead of something correct, and a model that papers over the gap without telling you. So that's where my time goes. I spend it on chunking and retrieval strategy, on evals you can rerun against every change, on getting the thing to admit it doesn't know instead of guessing, and on hard ceilings for latency and spend, because a retry loop nobody's watching is how API bills get out of hand. I also built the image generation and video-to-frames pipeline behind Dream Engine. For Docsum, a YC-backed contract intelligence platform, I built the admin layer that replaced hand-editing their production database every time they onboarded a customer. None of it holds up without the data layer underneath, which is where I spent the years before this. I was the founding engineer at Tajir, a YC W20 B2B commerce company, and later its tech lead running a team of four. I moved them off NoSQL onto SQL-backed APIs, then ran almost-zero-downtime schema migrations against a live operation that was processing real orders the entire time. I also wired Prophet forecasting models into transactional SQL workflows, which is where I learned how much of an ML problem is usually a data problem. Stack: Python, FastAPI, TypeScript, PostgreSQL, Supabase, and Claude Code every day. Tell me what you need the model to be right about and I'll tell you how I'd prove it is.