Technical voice on AWS pre-sales calls alongside Minfy’s sales team.
Translate client problems into proposed AWS architectures, cost estimates, and SOW scope for SMB and startup prospects.
Pitched Minfy’s GenAI and AWS migration work at Amazon HQ2, leading to follow-up prospect calls.
Authored the Statement of Work for an AI marketing-agent POC.
Bedrock Claude + Nova Canvas, deterministic HTML renderer, MCP server, and a human approval gate — 10 measurable objectives, 4 gated phases, 15 deliverables.
Build AWS cost estimates for proposals.
Model costs across volume tiers in the AWS Pricing Calculator and share estimate links with clients; POC budgets for IoT, GenAI marketing, and document-AI proposals.
AI Engineer — Jan 2026 – Present
2nd place, AWS × Anthropic Energy Hackathon (Houston, Sep 2026) — pipeline leak-detection agent on Amazon Bedrock AgentCore.
Built the detection and leak-localization tools and deployed the agent (Strands + Claude Sonnet 4.5): 5/5 leaks caught, 15/15 false alarms rejected, ~5-mile average localization on a 200-mile line.
Coordinated a three-engineer team through Claude Code with a shared CLAUDE.md git-sync protocol — scaffold to deployed agent in one build day.
Contributing to a personalization POC for Grubhub’s TensorFlow shared-bottom multi-task Topics Ranking model.
Fixed the Spark-on-EMR training-data pipeline for user-level behavioral features — a partition-read bug had silently zero-filled every order-history feature (0.36M rows read instead of 167.9M).
Owned preprod load and latency testing: 1.1M requests in a 3-hour test, up to 166 req/s (1.7× production’s peak) at 3.9 ms p95 model latency and 0.0002% errors.
Designed a cost-aware computer vision + VLM pipeline for Samsara dashcam crash footage.
Dynamic frame sampling and reduction, YOLO detections with ByteTrack tracking, bounding-box overlays, and structured JSON context before video summarization.
Defined a five-metric evaluation strategy (3 pipeline tracks × 4 VLMs on 97 labeled clips) with a YOLO baseline comparison.
Built an AI evaluation test bench that black-box tests any LLM endpoint — RAG chatbots, multi-turn assistants, and NL-to-SQL APIs.
Document and SQL sources of truth, custom question and ground-truth datasets, and multi-turn conversational testing.
Automated endpoint comparisons using semantic-similarity, rule-based, and LLM-as-a-judge scoring; ships as a FastAPI UI, CLI, and Docker image.
Built POS integrations and a Chat-to-BI assistant for a restaurant reservation-auction platform.
Square and Clover integrations (Toast designed) automating booking creation, customer mapping, order sync, and check-status updates.
Deterministic natural-language-to-BI workflow on Amazon Bedrock: intent routing into constrained SQL nodes with validated queries and auto-generated charts.