Five-plus years on the same stubborn question: how do you take something that works in a notebook and make it survive real users? Mostly that has meant backend systems, production machine learning, and the harness layer in between. Lately a fourth thread has crept in, machine learning for biology, where the datasets are small, the labels are noisy, and being honest about your numbers matters more than beating a leaderboard.
Since Aug 2026
An AI customer-experience platform used by telecom operators and several hundred smaller businesses. I work on the AI side: agents, retrieval, and the serving infrastructure underneath.
Since 2024
A small studio for applied AI that does not fit a product roadmap: sports science for elite cricket, clinical software for doctors, agriculture AI, and research engineering.
01
A model is only as useful as the surface you hand it. Most teams give an LLM raw tables and generated specs instead of the intent-named operations their own code already has.
02
Anti-inflammatory peptide prediction, and structure-aware virtual screening of natural products against understudied viral proteases. Both taught the same lesson: the dataset you choose decides the result long before the model does.
03
A bank rents the biggest MoE around, salary day brings a thousand customers, and the team learns what to prune, why Bangla breaks, and when to teach a student.
Mixture of Experts03
Five file conventions claim to be the standard for AI agents. Here is what each one answers, who really reads it, and the evidence behind both claims.
Retrieval and RAG08
Every retrieval method assumes clean structured text. Getting it out of a PDF is its own problem, and Docling is the 2026 answer worth knowing.
Retrieval and RAG07
How to run keyword, dense, fusion and reranking together on modest hardware: where the memory actually goes, and which famous constants are folklore.
Retrieval and RAG06
The scarce skill now is not writing code but knowing your own system: what you solved, how it works, and whether a new bug is even yours.
A gentle, playful look at k-nearest neighbours, the lazy way a computer labels something new by asking the dots closest to it.