Blog / Index
Notes on backend development, zero-server architecture, and running AI locally.
I replayed 40 real sessions of my coding agent and counted what happened: 1,304 tool calls, two retractions, four optimizations I declined to build, a 56% cut to my system prompt, and one feature that ran for days without telling me anything.
How I built skill-optimizer-gepa: an eval set from my real job applications, a scoring harness with hard checks and an LLM judge, and a GEPA run that plateaued, hallucinated a rule, and still ended 12.5 points above my hand-written skill. Every failure included.
Running local LLMs means private data stays on your machine, zero per-token costs, and full control over models and prompts. Here is why local models are my default for daily AI tooling.
The full story of building ejobtrack — a browser-only job tracker that parses Gmail with OAuth 2.0, regex parsers for 50+ ATS platforms, and on-device Transformers.js ML. Including the 429 rate-limit war, the two-pass sync bug, and the Google verification saga that ended in a paid annual security assessment I chose to decline.
As a DOST Region V backend intern, I led a 3-person team building REST API microservices in TypeScript and Node.js, refactored services from Clean Architecture to MVC, and integrated PostgreSQL, Docker, and Jest into an Agile workflow.