Blog

Writing & notes.

Lessons from shipping agentic AI, computer-vision, and MLOps systems to production.

May 26, 2026
Why a multi-agent system beats one big agent

A single do-everything agent conflates requirements, design, code, and tests into one fuzzy prompt. Splitting the work into roles makes the whole thing inspectable — and better.

LLM AgentsMulti-agentArchitecture
May 20, 2026
Building an eKYC pipeline that doesn't fall for a printed photo

How detection, liveness, and recognition fit together into a face-verification flow — and why the order of those three stages matters more than the models themselves.

Computer VisioneKYCFace Recognition