PRISM: why a similarity score isn't a probability
Finding one face across many cameras with ArcFace and FAISS, and the design choices that keep a human in the loop.
Short write-ups on what I built and what it taught me, grouped by the year the work happened. Architectures, flows and the bugs that taught me the most.
Finding one face across many cameras with ArcFace and FAISS, and the design choices that keep a human in the loop.
Select a year to see what I was working on and filter the write-ups below.
Hallucinations, brittle prompts and language gaps: the failure patterns I learned to look for while testing and ranking model responses.
Four input types, one card generator. Why I reduced everything to text first, and where that choice hurts.
EduInsight worked locally and broke on Netlify. A short list of what failed and what fixed it.
Tasks, chat, docs, an AI assistant and outreach automation in one Next.js app. The architecture and the bugs that taught us most.
A wellness companion for exam students handles sensitive text. Here is where the model is trusted, and where it isn't.
Finding one face across many cameras with ArcFace and FAISS, and the design choices that keep a human in the loop.
Written in 2026 as retrospectives on the year each piece of work happened.