Series
Multi-part deep dives into AI engineering projects, from concept to production.
An evolutionary forge where a cheap frozen LLM writes candidate code and a deterministic verifier scores it — reaching results that rivaled large-scale RL, on one desktop for a few dollars. From Atari Pong's theoretical ceiling (21-0 in all 40 games) to a cache eviction policy that beat twenty years of classics on production traces.
Measuring outcome per dollar for LLM agents — KOI (= KPI ÷ cost) — with the open-source tool Nishiki. From the VOI/KOI concept to hands-on measurement on real tasks like Text-to-SQL.
A hands-on verification of whether OpenClaw — a much-discussed AI agent platform — is actually usable and secure in production. Covers cost traps, model selection, agent runaway, and multi-agent design, for anyone considering adopting AI agents.
Migrating a virtual try-on system from PASTA-GAN++ to Gemini and Vertex AI — replacing an entire GPU pipeline with API calls and resolving the body diversity problem that GANs could never solve.
Building an ML-powered meal planning system to solve the daily struggle of nutritious, varied cooking — from cosine similarity to LSTM to ChatGPT.
Building a browser-based investment education game for children — from board game concept to Next.js + Phaser 3 production app with real Japanese economic history.