I build AI systems that work outside the lab — from edge-deployed computer vision on poultry farms to enterprise NLP pipelines at scale. I care as much about the engineers who keep those systems running as about the models inside them, and I write about what I learn along the way.
I care about systems that actually run — on constrained hardware, against messy data, in places far from the cloud. The interesting problems are the ones the demos never show.
Nothing published yet — so here's the running order instead. Each piece comes out of a system I've actually had to keep alive in production, which means they'll read as build notes, not think-pieces. The first one is being written now.
RTSP streams that drop at 3am, a Tailscale tunnel to a farm with bad connectivity, and 923 live birds that refuse to hold still for the camera. The gap between a notebook that hits good mAP and a box that survives a week unattended in a shed.
How concurrent LLM summarization, embedding caches, k-amplification and ef_search tuning took roughly 40% off a production retrieval pipeline.
What moved the needle taking engineers from notebooks to systems that run unattended — and the habits that mattered more than the code review.
Why research visibility compounds, what applied-science screens actually filter on, and how that looks from Hyderabad.
The first one lands on this page. If you'd like a nudge when it does — say so by email or follow along on LinkedIn.
I'm always open to thoughtful conversations — about AI, agritech, building careful systems, or anything in between.