Writing
Notes, mostly written to stop re-deriving things.
Infrequent. Usually the thing I wished existed while I was stuck on it. Older pieces also live on Medium.
TrustMesh: how we built a real-time fraud brain that thinks in graphs and sequences
A heterogeneous GNN chained into a Transformer, scoring Return-to-Origin risk on millions of daily Cash-on-Delivery orders inside a 50 ms p99 budget. Why graph topology and behavioural sequences turn out to catch different fraud, what the ablations showed on cold-start and clean-profile abusers, and what it took to serve it. Co-authored; accepted at AAAI 2026.
ReadDeploy a Flask app on Heroku using Docker
Containerising a Detectron2 model and shipping it. Everything I had to piece together from a dozen tutorials that were either too deep or plainly incomplete, written down once so nobody else loses the day to it.
ReadMore soon
Still owed: the ranking half of the story.
TrustMesh covers the trust side. The multi-task feed ranker — conflicting objectives, soft negatives, learning from skips — is the one I keep starting and not finishing. Until then the work page is the short version.