Vincent Li

Architected and built Aether, an end-to-end Safe GenAI Platform integrating four microservices: Atlas (traffic governance), Sentinel (safety supervision), Hyperion (ML inference), and MonitorX (observability). Implements defense-in-depth safety with tiered analysis (heuristics, ML classifiers, LLM-based detection) and quota-managed safety compute budgets.

Independently built Atlas, a production-oriented LLM gateway reference implementation with Redis-backed quotas, token-aware rate limiting, priority traffic shaping, dynamic stream reservations, and usage forecasting. Validated locally with 46 automated tests; not operated under sustained production traffic.

Independently built Sentinel, a production-oriented GenAI safety reference implementation with tiered analysis, PII detection/redaction, prompt-injection protection, and toxicity filtering. Previously served through RapidAPI and Cloud Run; validated locally rather than under sustained production traffic.

Contributing to core inference serving features and observability enhancements in one of the most widely adopted open-source LLM inference engines.

Independently built Hyperion, a production-oriented ML inference reference implementation covering GPU-aware serving, dynamic batching, Redis caching, Kubernetes autoscaling, Prometheus metrics, tracing, and alerting. A local audit passes 82 tests with 7 remaining failures; no production latency or throughput claims are made.

Independently built MonitorX, a production-oriented ML observability reference implementation with Python SDK instrumentation, drift and resource metrics, multi-channel alerting, FastAPI, InfluxDB, and Streamlit. All 97 non-API tests pass locally; the API test module currently has a stale import that blocks full-suite collection.

Authored a curated guidebook on LLM infrastructure, covering the full lifecycle — pre-training, post-training, inference, optimization, and monitoring. Includes references, diagrams, and structured learning paths.