Writing — Swayam Mishra

Every part was right. The answer was still wrong. — An AI affordability agent passed 797 tests and placed 48th of 3,062. Every component was correct. The output was still wrong, and the gap is the lesson.
Google DeepMind hackathon: a 59MB APK and a sticker sheet — I built an on-device Gemma agent for the Google DeepMind Bangalore hackathon, never got a live demo, and learned the printed rubric isn't always the real brief.
I ranked 8th and 577th with the same AI agent — One 24-hour agent scored four ways: 8th of 1,983 on the transcript stage and 577th on output. Same artifact, same day. The gap is the lesson.
How I Built a 15-Stage RAG Support Triage Agent in 24 Hours — A technical walkthrough of a support agent built for a hackathon: the architecture, the code, the bugs that cost me points, and what I would do differently.
How to build AI agents you can actually trust — A practical guide to reliable, auditable agents, using a vision claim-verifier (3rd of 1,773 at HackerRank Orchestrate) as the worked example.
Five hours, one cabin, and a hackathon build we kept — A payments API changes one word in its docs and quietly breaks your billing. We spent a Saturday in Bengaluru building the thing that catches it.
Where a Go API Stops Getting Faster — I rebuilt a Go URL shortener ten times. An index and a connection pool made it 106x faster. Then throughput stopped moving, and I went looking for the wall.