Research

My research has one throughline across a career move from academic HCI to industrial ML: making complex, high-capacity systems (deep networks, then LLMs) smaller, faster and more trustworthy without giving up the capability that makes them useful in the first place.

Current: efficient & reliable LLM/ML systems

2018 – present, Intuit AI

When does bypassing the language model help?

Multimodal speech-intent classification

A systematic architecture-search and fine-tuning study on real-time multimodal speech-intent classification. The finding: fusing lightweight, frozen audio and text encoders through a compact classifier reaches the production accuracy target (macro F1 = 0.746 ± 0.002) at roughly 1,000x fewer parameters than a frontier multimodal LLM pipeline, and a fraction of the latency — a counterintuitive result on when generative language-model layers help vs. hurt discriminative tasks. Write-up in preparation for external submission.

Fine-tuned small models as frontier-LLM replacements

Memory/context extraction, conversational coaching, model confidence scoring

Applied studies on replacing costly frontier-LLM calls with fine-tuned small language models in latency-sensitive production settings — using weight-space model merging and constrained decoding to recover reliability lost in the switch. One deployed example cut p95 serving latency roughly 8x while lifting structured-output reliability to 97%+. A related thread: building lightweight confidence/quality scorers that agree with an LLM-judge closely enough (0.80 Pearson / 0.88 AUROC) to gate production traffic without calling the judge model on every request.

Retrieval, recommendation & conversational agents

RAG systems, knowledge-base recommendation, GenAI assistants

Design and delivery of retrieval-augmented conversational agents and recommendation systems for customer support at scale — from an early domain-specific embedding/search model (built as a founding member of Intuit's first data-science team) to a proprietary RAG-based GenAI agent platform and the recommendation model behind Intuit's current GenAI assistant for support experts.

Earlier: information reliability & visual analytics

Ph.D. research, UC Santa Barbara (2011–2016), plus research stints at Yahoo Labs, Adobe Research, IBM T.J. Watson, and Nokia

My Ph.D. work asked a version of the same question I still ask today, just about people instead of models: how do you help someone tell reliable information from unreliable information in a noisy, high-volume stream (social media, in that case), and how do you design an interface that makes that judgment easier rather than harder? That work spanned:

See Projects → earlier work for specifics, and the Publications page for the papers.