I build AI products end-to-end — the research, the engineering, and the ship.

Each case study on this site is written in the same shape: situation, the decision I made, the option I killed, what I actually shipped, the one metric I'd watch, and what would tell me I was wrong. This is my product discovery, prioritization, and roadmap language — compressed into a page.

That shape is how I think about product: user problem first, scope the smallest shippable version, cut what won't earn its keep, name the KPI, name the failure signal. It's the same discipline whether I'm scoping a release, prioritizing a bug queue against an App Store deadline, designing a research benchmark, or picking an abstention threshold in a production LLM.

If you want to see how I'd approach a decision in your product — or a review in your research group — read any of the case studies below. The shape is the answer.

I'm Nidhi Pandya, an applied-AI engineer, product builder, and researcher based in New York City. My work sits at the intersection of NLP/LLMs and shipped product — from user research and product scoping through model fine-tuning, retrieval and grounding pipelines, all the way to production deployment. I own the full arc: discovery, prioritization, build, ship, measure.

On the research side I fine-tune large language models with LoRA/PEFT, build retrieval-augmented generation pipelines, and care a lot about trustworthy AI — grounding, citation, and knowing when a model should abstain instead of guess. My first-author paper (arXiv:2601.08852) shows an open-weight LLM matching commercial frontier models on claim-extraction accuracy for ~$15 of compute.

On the product and engineering side I took a full-stack civic platform (voiceapp.live) from a blank repo through production deployment and iOS App Store submission — solo. Every user problem, every layer, every deploy, every rollback story.

Before my master's, I spent 5+ years in enterprise software — as a functional / technical consultant deploying platforms into client environments in India, and in technical support before that. That's where I learned how product actually lands with the people who use it, and how requirements survive contact with reality. That mix — research depth + hands-on building + real delivery — is the lens I bring to everything.

I'm drawn to consequential, messy, real-world problems — civic technology, trustworthy AI, geospatial systems — where good product thinking and honest engineering both matter.