machine learning engineer bengaluru, in
Mostly trying to make small things work a little better than yesterday.
↓ scroll, slowly
01 — About
I build and evaluate LLM systems — retrieval pipelines, eval harnesses, and the MLOps around them. Most weeks: shipping models to production, then measuring what actually breaks.
02 — Work
Probe42 — Bengaluru
Productionising ML — moving models from Jupyter notebooks onto EC2 as reliable services with monitoring, retraining, and clean interfaces.
Built a BERT-based semantic search system and a generative model for synthetic compliance data, now used across internal tooling. Own the full path: data pipelines, training, evaluation, deployment, and the boring glue in between.
with Prof. Kripa Bandhu Ghosh — Kolkata
Fine-tuned a multilingual hate-speech classifier across several Indian languages on ~1.5M tweets.
Learned how much of applied LLM research is careful eval design, dataset curation, and reproducible pipelines.
IISER Bhopal
Econometrics and statistics give a strong mathematical foundation for the ML work — especially around causal inference, portfolio optimisation, and reading models honestly.
A habit, from economics, of not trusting clean-looking numbers.
04 — Stack
A small, boring stack —
chosen mostly because it gets out of the way.
05 — Writing
Notes I’ve written down — partly to remember, partly to be corrected.
Nominated by the program chair committee to review emerging ideas in sustainable development for health.
A small paper on Indian pension reform. Mostly accounting and assumptions.
10,000 episodes, ~9k lines of code, 43% win rate. Notes on shaping rewards for a game full of luck.
53.7% directional accuracy, and an honest write-up of where the model breaks and why markets don’t forgive it.