I build the systems that decide whether AI models and agents are ready to ship.

I evaluate voice and chat agents end to end at Microsoft: simulate real customers, score every conversation, and find out why agents fail.

01Focus

simulation

Simulated customers that call and chat with deployed agents, the way real customers do.

measurement

Metrics anchored to expected outcomes and grounding data, not vibes.

reliability

LLM judges that give the same answer twice, audited before they are trusted.

diagnosis

Tracing every failure back to why the agent failed.

2026 → now

AI Applied Scientist II · Microsoft

Building the evaluation systems for Dynamics 365 Contact Center AI agents: simulation, scoring and failure diagnosis.

2025

Data Scientist · Microsoft

Demand forecasting across Azure data centers; production runtime from 7 minutes to 1.

2023 → 2024

Applied AI Researcher · Duke University

Fine-tuned Llama-7B; built and evaluated a RAG course assistant with an LLM judge.

2022

ML Scientist · Workit Health

Patient chatbot with 1,000 weekly users; message classifiers deployed into the health record system.

🔒 private · in progress

fig. 01Can models write good evals?A mutation-testing benchmark for agent evals.

topology-ai

fig. 02Topology AINatural-language questions over network topology, checked against deterministic answers.

medredact

fig. 03MedRedactFine-tuned Mistral-7B that redacts medical notes.

protein-explorer

fig. 04Protein ExplorerDisease annotations predicted from ESM2 protein embeddings.