Retrieval · Evaluation · Agentic systems

Hey, I'm Shubham.

I work on making LLM systems trustworthy once they leave the notebook: retrieval that actually retrieves, and evaluations that measure what they claim to. Most of my research happens inside real products, with messy data, real users, and a pager.

You can find my work on my Research page, and find out more about me in general here. Or use these clickables to see where else I exist on the internet:

What I keep thinking about

01
Retrieval on the long tail. Hybrid lexical + dense search, rank fusion, and why rare queries are where systems quietly fail.
02
Evaluating the evaluator. Golden sets, LLM-as-judge calibration, and the provenance of the labels we grade against.
03
Agentic systems that fail gracefully. Multi-step orchestration, state, routing between models, and observability for each step.
04
Cheaper, more truthful models. Fine-tuning against hallucination, token-level cost work, and (from college) GPU/CUDA optimisation.