Jeffrey Chien

Contributed to model integration and serving workflows for Bloomberg GPT, including finance-specific sentiment analysis, entity recognition, and question answering. Built data preparation and evaluation pipelines covering 1500 benchmark examples to compare model quality across releases. Optimized GPU batching and scheduling, increasing inference throughput by 22% under comparable workload and hardware conditions. Added distributed tracing and GPU monitoring using OpenTelemetry, Datadog, and Grafana to support production diagnostics and capacity planning.

Built natural-language search and question-answering features using retrieval-augmented generation across financial data, news, and documents. Developed vector search/ranking/grounding/source attribution pipelines, improving retrieval recall@10 from 72% to 83% on an internal dataset. Designed prompt templates and guardrails with automated evaluations, reducing unsupported answers from 12% to 8% on a fixed test set. Integrated LLM services with entitlement checks and audit logging to enforce document access permissions throughout retrieval and answer generation.

Owned end-to-end development of back-end services and platform features for the Bloomberg Terminal. Optimized caching and request routing to reduce p95 API latency (250 ms to 180 ms) for targeted endpoints. Built release automation, feature flags, and regression tests, increasing release frequency to three times per week. Implemented monitoring, alerting, and incident runbooks, reducing average recovery time from 50 minutes to 35 minutes.