Experience

Mountain View, CA · Feed AI — LLM for RecSys & post-training

• Contributing to LinkedIn's out-of-network retrieval for homepage-feed candidate generation, applying LLM-based generative retrieval to surface relevant content beyond a member's immediate network.
• Led a research direction on adaptive, compressed member representation — modeling profile signals, temporal engagement history, long-history compression, and member-regime differences in retrieval.
• Built a multi-axis retrieval experimentation framework studying tradeoffs among Recall@K, diversity, efficiency, and evaluation fidelity across cold-start, sparse-, stale-, and active-user scenarios.
• Developed a reusable out-of-network experimentation pipeline (training, retrieval eval, config reuse, execution tracking) improving team velocity and reproducibility.

Efficient & Intelligent Computing Lab (NVIDIA) · Atlanta, GA · LLM evaluation & agentic AI

• Co-authored AgentSuite (ICML 2026), an automated pipeline auditing validity flaws in LLM-agent benchmarks; matched expert judgments at 0.79–0.87 F1 across 6 benchmarks and showed flaws affect ~23% of tasks, reshuffling 63% of model rankings on a 30-LLM leaderboard.
• Built the taxonomy-guided LLM-judge stage and rule-based structural detectors, decomposing agent tasks into user / environment / ground-truth / evaluation components; generalized to unseen benchmarks with no retuning and surfaced 100+ undocumented issues.

Graph Mining Lab · Atlanta, GA · AI for science, graph representation learning

• First-authored a causal brain-connectivity method (AIME 2025) recovering directional dependencies invisible to correlation graphs, improving diagnostic prediction across ~8K patients via Granger-directed graphs.
• Co-authored a state-based Transformer (IEEE EMBC 2025, oral) modeling temporal transitions across latent functional-connectivity states, outperforming static-connectivity baselines on clinical prediction.