case studies
How they work, and what broke
Architecture, design decisions and the bugs that cost real time, for the projects where I kept the notes.
Reinforcement learning
PokeGen3-RL
RL agent that plays Gen 3 Pokémon battles by reading emulator memory directly. A Lua bridge inside mGBA drives the game; a Gymnasium env handles masking, rewards and an overworld autopilot.
8 sections · 10 deep-dive write-ups
Full-stack platform
TraceVault
Knowledge platform for engineering experience: write up bugs, trade-offs and design calls, then search, share, save and discuss them. Feed, vault, collections, live threads and email MFA.
7 sections · 5 deep-dive write-ups
AI · Hackathon
Setu
Multilingual platform that turns citizen requests into ranked, explainable district priorities. Batched Gemini extraction with a validation harness, fused with census data.
6 sections · 6 deep-dive write-ups
AI assistant
Academigo
Study companion with a React + HeroUI front end, a small Django API that keeps the Langflow token server-side, a Botpress chat surface, browser-side history and light and dark themes.
7 sections · 5 deep-dive write-ups
Developer tooling
ACE
Terminal assistant for developers: a project registry with instant navigation, a guarded Git save flow, scheduling, backups, a tmux dashboard and a Textual TUI. Installable with pipx, tested on Python 3.10 to 3.14.
8 sections · 6 deep-dive write-ups