Software Engineer at Google · AI Infrastructure

I build tools for understanding how AI systems behave under load—and what happens when they fail. My focus is the production layer around models: inference serving, request routing, observability, and reliability.

My foundation is network and infrastructure engineering: routing, load balancing, traffic engineering, and operational debugging. I bring that perspective to LLM systems, where a successful HTTP response can still hide a slow first token, a stalled stream, or an ineffective fallback.

From latency symptoms to an evidence-backed incident report.

An investigation agent that combines serving telemetry, ranked hypotheses, diagnostic experiments, and explicit uncertainty. Includes Google ADK tools, read-only vLLM metrics collection, persistent investigation history, and a dashboard. A seeded simulator makes incident scenarios reproducible without a GPU.

Engineering focus: diagnostic reasoning, telemetry contracts, and controlled comparisons. Simulator validation is distinct from live GPU-backed serving validation.

Turn streaming behavior into an operational verdict.

Checks streaming endpoint behavior against configured service-level objectives, including time to first token and request outcomes. Produces PASS, FAIL, or INCONCLUSIVE verdicts for CI workflows, keeping insufficient evidence separate from a healthy result.

Engineering focus: meaningful reliability checks, configurable thresholds, and actionable failure reporting.

Explore the mechanisms behind serving performance.

A reproducible lab for scheduling, batching, KV-cache pressure, and the relationship between throughput and useful work delivered within latency targets. Pairs GPU-free simulations with a vLLM Colab companion for experiments against a real serving engine.

Engineering focus: workload design and performance trade-offs. Simulated results and measured serving results are kept distinct.

Make failure handling observable and testable.

A gateway sandbox for retries, deadlines, fallback routing, circuit breakers, and rate limiting. Mock services, load tests, and Prometheus metrics make it possible to examine how resilience controls interact under failure.

Engineering focus: bounded failure handling and overload behavior. The lab uses simulated services rather than production model traffic.

I contribute fixes and tests to the systems around inference. A few examples:

Merged

• LiteLLM #43959 — Resume synchronous streaming fallbacks without retrying the primary deployment.
• Agent Router #2299 — Add OpenRouter real-provider data-plane test coverage.

Under review · status checked October 17, 2026

• vLLM #59439 — Propagate SIGTERM during engine startup.
• Triton Server #8987 — Preserve gRPC trace metadata length.
• OpenTelemetry #640 — Improve LangChain token-usage fallbacks.

Streaming latency · batching and KV-cache behavior · routing and fallbacks · overload protection · tracing and SLOs

For a smaller monitoring tool, see LLM SLO Probe: black-box streaming measurements and Prometheus textfile output.

I welcome technical discussion through issues and pull requests on these repositories—especially reproducible failures, measurement design, and serving reliability.