My through-line is proxy fidelity across the ML pipeline. Training data is a proxy for the real distribution; a model's learned objective is a proxy for what you actually wanted; an evaluation metric is a proxy for whether it's genuinely good and safe. The same gap shows up at all three — and it compounds: a proxy-corrupted dataset can train a proxy-gaming model that a proxy-blind eval scores as clean. I study those failures at the interfaces, across the whole chain, not inside any single stage. Where the gap shows up:
- Data — is a benchmark a faithful proxy for real task quality? (AgentSuite, ICML 2026)
- Model — did it learn the objective we meant, or one it can game? (retrieval & recommender systems)
- Eval — does a metric or safety monitor actually track quality and safety? (AgentSuite's LLM-judge, vs. expert labels)
I'm finishing an MS in Computer Science at Georgia Tech (Dec 2026) and spending the summer as an ML engineer intern at LinkedIn (Feed AI). Across all of it, the same question keeps recurring — whether the numbers we report mean what we claim they mean.