Byungkyu (Jay) Kang
I build and ship production machine-learning and language-model systems — from efficient speech/text classifiers to fine-tuned LLM-based agents — currently as a Senior Staff AI Scientist at Intuit, where my models run in production for 50M+ consumers. My background is in HCI-flavored AI research (Ph.D., UC Santa Barbara), and I still care most about systems that are fast, interpretable, and reliable enough to trust in production, not just accurate on a benchmark.
What I work on
A quick tour — see Projects for the full writeups.
Intuit Assist — GenAI assistant for support experts
Led the company's GenAI conversational assistant used by customer-support experts, plus a real-time recommendation model that routes customers to the right support channel.
Replacing frontier LLMs with lightweight fine-tuned models
A recurring theme in my recent work: showing that a costly frontier LLM call can often be replaced by a small, fine-tuned model at a fraction of the latency and cost — without giving up accuracy. Three shipped examples: a real-time speech-intent classifier (macro F1 0.746 at ~1,000x fewer parameters), a memory/context-extraction model (~8x lower latency, 97%+ reliability), and a live sales-coaching model.
Production ML infrastructure
Own training/eval infrastructure across cloud ML platforms, including a paved-path fine-tuning/RL library with an automated-research-agent feature for running architecture and hyperparameter searches unattended.
Background
Before Intuit, I was a data scientist at Nokia Technologies (analytics for device/content signals) and a visiting scholar at IBM's T.J. Watson Research Center (graph mining and interactive visual analytics). My Ph.D. work at UC Santa Barbara focused on information reliability on the social web and intelligent, visualization-driven user interfaces — an interest in making complex systems inspectable and trustworthy that still runs through my current work on production LLM systems.