Hands-on engineering leader building dependable AI-agent systems, evaluation infrastructure, and security tooling.
I turn ambiguous ideas into tested, operable products—from architecture and agent workflows to full-stack implementation, security boundaries, release engineering, and contributor experience. I maintain SAF-MCP, contribute to Promptfoo, and am the lead author of ConformalGuard, published at COLM 2026. Previously at American Express and Wells Fargo.
I am especially interested in startup conversations where technical leadership, product ownership, and hands-on founding-engineer execution matter.
- Technical leadership: architecture, 0→1 product engineering, platform design, full-stack delivery, release operations, and open-source maintenance
- AI systems: coding agents, LangGraph and multi-agent workflows, LLM evaluation and red teaming, human-approval boundaries, retrieval, and local-first AI
- Real-time AI: voice pipelines with Pipecat, Deepgram, WebRTC/WebSockets, streaming interfaces, and structured function calling
- Security: Model Context Protocol (MCP), threat modeling, detection engineering, least privilege, auditability, and privacy-preserving system design
- Core stack: Python, TypeScript, LangGraph, React, FastAPI, Hono, Cloudflare Workers/D1/R2, SQLite, Swift/SwiftUI, and agent-development tooling
- Based in: San Francisco Bay Area
SAF-MCP — Maintainer
A MITRE ATT&CK-style framework for documenting and mitigating threats in Model Context Protocol and agentic systems. My contributions span threat research, Sigma detection rules, security-control design, catalog validation, and contributor infrastructure.
Selected work:
- Documented and tested detection for MCP Inspector exploitation, path traversal through file tools, and AI agent CLI weaponization
- Designed out-of-band authorization, anomaly detection, query guardrails, and output DLP for privileged tool invocations
- Strengthened catalog integrity through detection-rule coverage, MITRE mapping corrections, and mitigation cross-references
- Built contributor validation and approval-gated onboarding workflows for SAF Agentic Use Case Analysis
Promptfoo — Contributor
Merged contributions to the open-source platform for testing and red-teaming prompts, agents, and RAG systems:
- Made
afterEachevaluation-hook metadata and named scores flow through to database persistence, file outputs, metrics, and the web viewer - Fixed OpenAI Codex SDK cached-token and cost accounting
- Added CLI list-command coverage and provider-validation edge-case tests
Additional upstream work includes a benchmark-backed audit-log read-replica fix in Pangolin, a Python library behavior fix in Supabase, and module caching in the American Express One App platform.
Lead author. Published at the Conference on Language Modeling (COLM) 2026.
ConformalGuard is a model-agnostic calibration layer for choosing pre-deployment action thresholds with finite-sample false-positive control under exchangeability. We evaluated it with fine-tuned BERT, Claude Sonnet 4, and GPT-4o on public email corpora, exposing the recall and robustness trade-offs of coarse zero-shot LM scores.
arXiv paper · Code and reproducibility · Paper PDF · OpenReview record
| Project | Technical focus |
|---|---|
| OpenBird | Open-source, local-first personal AI memory for macOS. Text-first capture, transient on-device OCR, SQLite/FTS5/vector retrieval, cited answers, Ollama/BYO-model support, and bounded read-only MCP access. |
| ConfPilot | Open-source conference program operations from CFP to published schedule. Role-scoped workflows, idempotent lifecycle transitions, React/Hono, and runtime ports for Cloudflare D1/R2 or Node/SQLite. |
| CyclePel | Native SwiftUI indoor-cycling platform with FTMS smart-trainer control, structured ERG workouts, route simulation, and hardware safety gates. Currently private and being prepared for a future open-source release. |
| cc-statusline | A focused Claude Code statusline for model, directory, Git branch, and output-style context. |
| PlanGate | PLAN.md-driven development with explicit quality gates for Claude Code workflows. |
- Agent-native delivery: decompose work into bounded workstreams, use coding agents for implementation and review, and require exact verification before integration
- Dependability by design: deterministic evaluations, human approval for consequential actions, audit trails, privacy boundaries, and explicit failure states
- Full-lifecycle ownership: product definition, system architecture, implementation, testing, packaging, release operations, documentation, and contributor workflows
- Leader who still builds: stay close to code and users while creating the technical systems that let a team move quickly without losing correctness
- Organize Bay Area events through The AI Agents Community
- Teach workshops on Claude Code, Cursor, and AI-powered software development
- Build and contribute in public across agent security, evaluations, local-first AI, and developer infrastructure
If you are building coding agents, agent infrastructure, AI security, evaluation systems, or ambitious developer products—and need a hands-on technical leader who can move from architecture to shipped systems—let's connect.
Outside software, I run marathons and train for triathlons.