Selected work across LLM measurement, data infrastructure, online experiments, realtime agents, and research software. Each case study records the constraints, design decisions, and failures behind the implementation.
The pipeline recovered 30.1M archived pages, extracted structured measurements for 88,336 startups, built human gold standards, optimized LLM classifiers with DSPy/GEPA, and organized embeddings into 265 interpretable themes.
A six-condition platform that controls the evidence layer while varying AI behavior. Includes provenance-aware retrieval, structured outputs, SSE streaming, experiment assignment, Postgres persistence, and a research dashboard.
FastAPI and Google ADK backend with function-calling care tools, Firestore sessions, Vertex AI Memory Bank, plus a Twilio-to-Gemini Live audio path with transcoding, backpressure, barge-in, and reconnect handling.
Diarized transcription, LLM paragraph segmentation, source search, and structured evidence review for an HBS research team conducting an ongoing 331-interview program, including a localized Chinese-startup workflow.
Balanced assignment, reservations, resumable sessions, quality checks, and exports for human coding. The resulting 800+ annotations support inter-rater reliability and gold-label inputs to model evaluation.
These earlier projects cover controlled experiments, computational analysis, and human evaluation of AI systems.