Custom AI ModelsNotes from building
the whole system.
Production AI architecture from the person who shipped it: RAG, agentic workflows, real-time decision systems, and what changes when one architect owns a system end to end instead of one piece of it.
RAG isn't a feature. It's an architecture.
Why a production retrieval system lives or dies on the telemetry and evaluation wrapped around it, not the prompt at its center.
Read
What replacing four spreadsheets actually took
Building a company's AI infrastructure from zero means the hard part was never the model. It was registration, retention, and who's allowed to see what.
Read
Listening to a room in real time
A production audio-intelligence platform runs on 100+ DSP features and a continuous evaluation loop, not a clever model alone. Most of the engineering is in the loop.
Read
From statistics to LLMs: the same job, different tools
Eight and a half years across biostatistics, deep learning, and agentic systems taught one thing: the tools changed four times. The job never did.
Read