Independent AI architecture practice. Production multi-agent orchestration, autonomous AI agents, and consumer AI with layered guardrails. IEEE service in AI readiness for critical infrastructure.
Licensed P.E. · IEEE PES SC-5 (AI) Co-Chair · 22 U.S. Patents (9 pending; 2 filed 2026 in AI Agent Safety) · CISSP · CCSP
LangGraph pipelines coordinating concurrent agents across vendors. Cross-vendor adversarial verification catches hallucinations that single-model review misses.
Deterministic first-contact architectures, audit-queue gating on AI-proposed actions, state-machine escalation triggers, and operator-takeover flags. Safe-by-design.
AI architecture review boards, NIST AI RMF, ISO/IEC 42001, EU AI Act readiness. Authority anchored in IEEE-SA standards work on grid readiness for data center deployment.
Two-layer guardrail systems, tiered model routing, adversarial test coverage across prompt injection, jailbreak, and homoglyph evasion. Privacy-constrained observability.
Publicly verifiable at github.com/martymcenroe.
Five-stage LangGraph StateGraph pipeline (Triage→Design→Spec→TDD→PR) coordinating concurrent Claude and Gemini agents. Cross-vendor adversarial verification — Claude generates, Gemini reviews. Five-layer evaluation framework: execution-based verification, AST structural analysis, cross-model review, stagnation detection, longitudinal learning. Python, LangGraph 1.0, DynamoDB, GitHub Actions.
Production autonomous agent for recruiter triage. Architectural safety: deterministic first-contact template — zero AI on the highest-stakes message. Audit queue gating AI-proposed actions, 10-state conversation state machine with escalation triggers (legal threat, message-count overflow, VIP allowlist), and operator-takeover flag. Hybrid-cloud boundary at SQS: Cloudflare Worker ingests + pre-retrieves RAG, AWS Lambda processes. Two-stage Bedrock pipeline (Haiku for intent classification, Sonnet for generation) on returning conversations. Pipeline of quality gates: anti-hallucination, sensitive-data detection, AI-disclosure prevention, hostile-tone blocklist.
Seven-stage handler pipeline for consumer-facing AI with independent short-circuiting at each stage. Two-layer guardrails: deterministic denylist plus semantic LLM classifier with five-category taxonomy. Tiered routing — Haiku as default classifier, Opus invoked only as verifier when Haiku flags candidate prompt-injection. Privacy-constrained observability. AWS Lambda, Bedrock, DynamoDB, CloudFront, Chrome and Firefox extensions.
AI governance study platform for IAPP AIGP preparation. Built on the same orchestration and guardrail patterns as the production platforms.
Methodology and tooling for AI-assisted technical authorship at institutional scale. Forces substantive human contribution at every step. Anchored in USPTO Inventorship Guidance and Pannu v. Iolab. PolyForm Noncommercial 1.0.0.
Chrome extension for extracting LLM conversations to structured JSON. Preserves message ordering, role, code blocks, and tool-use markers across ChatGPT, Claude, and Gemini.
Active applied research on agent-context permission bits as an alternative to the "agent inherits user permissions" model that creates excessive blast radius for autonomous agents. Manifest specification, capability authorization design, filesystem-level prototype.
Independent practice. Engaged via C2C contract.
recruit@martymcenroe.ai