Four systems, end to end: from the index, to the agents that use it, to the evals that keep them honest.
End-to-end retrieval and ranking. A two-tower model learns dense user and item embeddings, FAISS handles candidate generation, and a PyTorch MLP reranks the shortlist. Beats matrix factorization and GBDT baselines on Recall@50, NDCG@10, and MRR.
An eight-node LangGraph agent that takes a deployment spec, generates adversarial attacks, judges every response against a seven-verdict rubric, and clusters failures into families with mitigations. It treats over-refusal as a first-class failure, not just harmful compliance.
A ReAct agent over three domain-partitioned corpora (rules, tactics, game situations) with hybrid dense and sparse retrieval plus cross-encoder reranking, on the hardest tier of SportQA. Retrieval precision was 99.8%, and 91 of 96 wrong answers still had the right evidence retrieved, so the ceiling was reasoning calibration rather than knowledge.
Team project with Cady He and Siyani Vengatagiri.
Scaled dot-product and grouped-query attention, RoPE, SwiGLU, LayerNorm, top-p sampling, and AdamW, all implemented from scratch in PyTorch. Training and eval pipelines cover text generation, zero-shot sentiment on SST and CFIMDB, and arithmetic reasoning.
Peer-reviewed and preprint.
Steven A. Lehr, , Mahzarin R. Banaji
PNAS Nexus, 2026
Built per-account personalization for the production recommender behind TI.com using a Thompson sampling contextual bandit with three-tier fallback. Shipped a new email recommendation channel to production with a FastAPI serving endpoint, drift-gated testing, and MLflow deployment. Designed the RAG subsystem for an on-prem agent that auto-generates SPICE testbenches for analog circuits.
My undergraduate thesis, conducted at the lab for BITS credit. Led a responsible AI evaluation of large multimodal models, auditing face-to-character inference biases across four frontier models. Built the preprocessing, bias-quantification, and multi-model evaluation pipelines behind 8,160 trials.
Designed synthetic data generation pipelines producing over 100,000 samples using GANs, VAEs, and kernel density estimation. Evaluated 15+ open-source LLMs on synthetic enterprise benchmarks with ROUGE, BERTScore, and COMET to guide internal model selection.
Built and deployed a conversational AI agent with LangChain and the OpenAI API for a US client, enabling real-time auction data analysis. Delivered the full stack with Flask, PostgreSQL, and React, and presented the architecture and a live demo to the CEO, CTO, and COO.
I've been captivated by ML and LLM agents since my junior year of undergrad and I have not really stopped thinking about it since. Most of what I build sits where retrieval meets evaluation: systems that find the right thing, and the harnesses that catch them when they don't. I care about that second half more than most people seem to. There are two sides to this technology, and the capability side already gets plenty of attention.
Outside of that: basketball, badminton, cricket, and an unreasonable number of movies. I like the ones that burn slow for an hour and then pull the rug out. I rapped in college. I do not anymore.
Open to new grad AI/ML engineering roles starting May 2027.