Hi, I'm Tanmay Somani

I build applied ML systems end to end: XGBoost baselines validated with time-series cross-validation before any deep model gets a look in, LSTM forecasts for Delhi's hourly air quality, Zhang et al.'s ECCV'16 colorization network running on CUDA, and a Double DQN agent trained on hand-engineered danger-state features. On the LLM side I've shipped local-first RAG pipelines — ChromaDB memory, Ollama inference, BERTopic + HDBSCAN for theme discovery.

Most of the time goes to the unglamorous parts that make models trustworthy: leakage-proof splits, error analysis before architecture tuning, property-based tests on financial math (pytest + Hypothesis), and wrapping the result behind FastAPI services people actually use.

>>> now: comparing seq2seq LSTM vs. lagged-gradient-boosting heads for multi-step AQI forecasting

How I work

Every new dataset gets the same treatment: profile it and plot the ugly parts first; score a dumb baseline under a validation scheme that mirrors deployment (time-aware for anything temporal); then earn complexity incrementally — error analysis decides where the next hour goes, not model fashion. Anything worth keeping ships behind an API, with tests on the math and eyes on the residuals.