Hey, I'm Nikhil 👋
I build AI systems that have to work on messy, real-world text, then I try to measure whether they actually do.
This fall I'm heading back to UNH for a PhD 🎓
Evals, specifically: evaluating agents around knowledge bases, RAG, and memory. Did the agent do the thing you meant, or did it just follow a script?
Before that I founded ResAlign AI, a career copilot. Not a mass-apply machine. More like: are you actually ready for this role, and is it even a fit? Alignment over spray-and-pray.
The research bug started as a class project during my master's. It turned into a paper on claim extraction for fact-checkers, published at CheckThat! (CLEF 2025). Same lesson as always: a high score isn't the same as being useful.
I came up through Electronics & Communication Engineering, IoT, and a detour through SAP security at TCS. Then a Masters in CS at UNH (May 2025). Never was a LeetCode person. More of a first-principles, breadth over depth and top-down thinking one 🛠️
### Now 🎓
Heading back to UNH this fall for a PhD on agent evals in knowledge bases, RAG, and memory systems. Scoring behavior against intent, not a rigid workflow.
- ResAlign AI: Career copilot. Fit, preparedness, alignment. Not another apply-to-everything button.
- HarnessBox: Sandbox + harness primitives for coding agents. Zero dependencies, on purpose.
- agentic-rag: Multimodal agentic RAG with a real eval harness. "It felt pretty good" is not a metric 🙂
- CheckThat: Noisy social posts → concise, checkable claims. Class project → paper →
pip install.
UNH at CheckThat! 2025: Fine-tuning Vs Prompting in Claim Extraction
Started as a class project. Ended up at CheckThat! (CLEF 2025).
We tried a bunch of fine-tuning and prompting setups to pull one succinct, checkable claim out of a messy social post. FLAN-T5 won on METEOR. Other methods like my iterative self-refinement loop sometimes produced claims a fact-checker would actually want to work with. That gap is the whole plot. Interesting to see Recursive Language Models growing popular today. Iterative refinement, Recursive Learning, Continual Learning, etc., These are all interesting things to keep your eye on if you're interested in making models better with data from real-environments.
Earlier: Teaching Neural nets to rationalize sentiment with human annotations and LIME. Can the model show its work in a way a person would recognize?
- MS, Computer Science @ UNH, 2022–2025
- IT Help Desk @ UNH, 2024–2025. Unglamorous. Made me faster at finding the actual failure.
- SAP Security Consultant @ TCS, 2021–2022. Roles, Authorization, SoD, the boring-but-load-bearing layer.
Knowledge bases, graphs, memory: what an agent actually knows, not just what it retrieved once.
Agent evals: score the behavior against intent, not a storyboarded workflow. A golden trace is a useful fixture. It is a terrible definition of success.
Inference engineering, lately a bit of an obsession. Disaggregated prefill/decode, speculative decoding, prompt caching, prefix caching. Making serving cheaper and faster without pretending that's the same as making the model smarter. KV cache, continuous batching, the whole rabbit hole.
Always happy to talk about research, agents, or a weird eval problem. LinkedIn or X is the easiest ping · site has the longer version.