Software engineer at AWS on Amazon Bedrock Guardrails. I build and operate production distributed systems at scale, and increasingly ship end-to-end by directing AI coding agents: automations, internal agents, and full-stack features.
Open to backend and AI-native engineering roles · Based in Bellevue, WA · Happy to chat over coffee.
Usually reply within a day.
Built end-to-end by directing AI coding agents from requirements through deployment. The tool supports on-call engineers in evaluating and executing customer quota increase requests, handling the decision context and the execution so the human just verifies and approves. Saves roughly 30 minutes per ticket across 100+ tickets a week, on the order of 50+ engineer-hours weekly.
When a new production ticket comes in, an agent is dispatched to gather CloudWatch faults and the relevant queries, correlate the logs toward a likely root cause, and compile customer-impact data after the incident. It compresses the most critical window, the first 30 minutes, so on-call can verify the findings and move straight to mitigation instead of assembling the picture by hand.
Created and maintained the scaffolding that makes the whole team more effective with AI agents: knowledge bases and context docs, steering and config files (CLAUDE.md, Kiro steering), reusable skills and hooks, and documented prompt patterns. The goal is consistency, so good results don't depend on who's driving the agent that day.
Delivered two new Bedrock Guardrails capabilities full-stack: applying safety policies to user prompts (previously only model outputs were checked) and a monitor-only mode that detects violations without blocking the request. Extended the public API schema, request validation, and DynamoDB data-access layer. Updated all of the existing policy-management UIs in React, and built backward-compatible logic so guardrails created before the launch continued to behave identically.
Primary engineer driving technical launch readiness for Bedrock Guardrails' first support for code as a content type, the gap that had to close as coding agents took off. Implemented the input and output transformations that feed code into the underlying safety model and translate its responses back into the public API's format. Coordinated across data science, product, documentation, and application-security teams and successfully delivered a re:Invent launch across 40+ regions.
Rolled out an in-memory caching layer across all production regions of a service handling 148M+ requests/day. Validated to 4,000 transactions/sec from a single account, eliminating a class of DynamoDB hot-partition throttling the service had no other way to mitigate. Production results: 53.5% reduction in DynamoDB read capacity on the metadata table, and up to 213 ms (-23.5%) drops at 99th-percentile latency for high-throughput customers. The hard part was consistency, not the cache: a dual-read pattern between two components of the same service created a race window where caching metadata and policy independently could surface stale authorization data. Chose a consolidated cache that's consistent by design, behind a feature flag.
An early AI project (2022), before LLMs went mainstream, exploring language models for music: fine-tuned GPT-2 on a structured music dataset encoded as token sequences, building the dataset preparation, tokenization, and training pipeline in Python. Adapted from Ens & Pasquier (2020).