Nick Ung

Nick Ung is a data science manager at Lyft helping develop Lyft AI Assist, the company’s automated customer-support platform for riders and drivers. His work focuses on ensuring agents can handle frustrated, skeptical, unpredictable customers before they enter production.

Ung interned at Lyft and ChowNow before returning to Lyft as a full-time data scientist. His account of his ChowNow internship describes work spanning algorithms, product analytics, and collaboration with business leaders. At Lyft, his career has included applied science, technical leadership, fraud, safety, and customer care. Working with Akshay Sharma, he has helped guide the direction of a self-service customer-support platform built around specialized agents, domain-expert configuration, and production monitoring.

How he makes support agents more dependable

• Offline evaluation as a launch gate: Agents must clear explicit acceptance criteria before deployment; live riders and drivers should not become unwitting test subjects.

• Production-representative user simulation: Conventional language models produce customers who are unrealistically polite and cooperative. Ung fine-tuned simulated users on real support interactions, creating more demanding tests that better reflect impatient riders, refund seekers, and skeptics of automated support.

• Multi-turn agent evaluation: Inspired by τ-bench, his approach tests LangGraph agents across complete simulated conversations, combining language-model judges with deterministic checks on whether the agent took required actions.

• Model, context, and harness learning: Evaluation failures can point to different remedies: model training, better information and retrieved knowledge, or changes to prompts, tools, routing, and orchestration.

Ung is also developing a configuration-driven evaluation harness that makes datasets, personas, tasks, and evaluators reusable across local development, pre-commit checks, and release pipelines. His work on practical agent evaluation extends toward fine-tuning and reward modeling informed by real customer interactions.