Research & Applied — Amazon
LLM Data Access Pipeline — Amazon
Owned end-to-end — problem definition, build, stakeholder alignment, adoption program, and outcome measurement. Built an LLM classification pipeline on AWS Bedrock to analyze security access request tickets — Python API to extract ticket data, stored in S3, classified in Jupyter Notebook using prompt engineering to identify what data customers were requesting and why. Used findings to: build new features in the Shepherd security product, create a new data access policy published internally, design an optimized data access request workflow, and reduce engineering oncall burden. Ran weekly office hours, trained the data team oncall, personally took ownership of access approvals during transition. Resolution time: 7 weeks → 1 week.
AWS Bedrock
Python
S3
Jupyter
Prompt Engineering
Shepherd Security
Trust by Design — Live Demo
Enterprise AI Risk Assessment Tool
Describe any AI initiative your organization is considering. Get an instant risk tier, architecture recommendation, governance requirements, regulatory flags, AI FinOps cost estimate, and 90-day implementation roadmap — in 30 seconds. Built on Claude API with enterprise AI frameworks encoded: NIST AI RMF, OWASP LLM Top 10, SOX, GDPR, ISO 27001. This is what I built manually at Google for 50+ AI programs over two years. Now it runs in 30 seconds.
Claude API
NIST AI RMF
OWASP LLM Top 10
Agentic AI
AI FinOps
Enterprise Architecture
Trust by Design — Live
AI Governance Intake Classifier
Built an AI-augmented governance intake system that takes a plain-text business use case, classifies risk tier (Low / Medium / High / Critical) using the Trust by Design framework, generates tailored discovery questions, and routes to the right approver — all in under 30 seconds. Architected on n8n's native AI Agent node (LangChain-based), with the risk policy pulled dynamically from GitHub at runtime for clean separation of concerns. Refactored from a brittle 13-node HTTP request pattern to a stable 4-node native agent architecture. Eliminates manual intake triage that previously took days.
n8n
Claude API
LangChain
Prompt Engineering
GitHub
In Progress
Agentic AI Research Workflow
Multi-step research agent that takes a governance question, retrieves relevant information from a knowledge base, synthesizes findings, and returns a structured recommendation with sources and confidence level.
LangGraph
Claude API
RAG
Vector DB
→ Live demo coming soon
Google — AI/ML Strategy
Enterprise AI Portfolio — Google SteerCo
Built and ran the operating model for Google Corporate Engineering's AI/ML program — intake framework, financial impact model, and prioritization framework across 50+ programs and 14 verticals. Led monthly SteerCo with 20+ Directors and SVPs. Portfolio represented $30–50M in AI investment.
Portfolio Management
LLM Governance
Executive Reporting
