Graham Anderson

ABOUT

Principal AI Engineer and AI Architect building verifiable agent systems, knowledge-graph platforms, and AI infrastructure for regulated technical environments. Most of my recent work sits in export-controlled aerospace and software-assurance settings, so I keep public claims focused on architecture, code, evaluation, interfaces, and technical briefing.

My path has been unusual: interactive production, data science, and now defense R&D. Earlier on, that included commercial and interactive work around familiar names like Sony, Adidas, X-Games, Disney, Microsoft, and Toyota; more recently it has meant aerospace, software assurance, and agent systems. I work best when the problem crosses disciplines and still needs one person to take responsibility for the complete result. Open to scoped consulting and selected Principal, Staff, Architect, LLM Platform, and Security/Compliance AI roles.

SELECTED IMPACT

• DARPA ARCOS / ACERT — technical lead for knowledge-graph and LLM-assisted certification of mission-critical software, briefing prime-contractor teams and program collaborators while remaining hands-on with architecture and code.
• Tau and agent-skills — built receipt-gated agent orchestration plus 340+ reusable agent capabilities and 90+ bounded worker roles, with deterministic checks on roughly 85% of the public capability set.
• Regulated graph-memory and compliance systems — built private RAG/GraphRAG systems over NIST, CWE, MITRE ATT&CK, D3FEND, SPARTA, and requirements-to-proof corpora; public details stay bounded because much client work is export-controlled.

EXPERIENCE

EDUCATION

• Metis Data Science Fellowship | 2016
• Trinity University | BS, Finance, Marketing, and Economics

CORE COMPETENCIES

Evals & Quality

LLM Evaluation, Agentic Evaluation Harnesses, Adversarial/Blind Testing, Regression Gates, Ground-Truth Fixtures

Observability & LLMOps

AI Observability, Drift Detection, LLMOps

Agentic Orchestration

Multi-Agent Systems, AI Agents, DAG Contracts, State Tracking, Tool Calling, Model Context Protocol (MCP), Bounded Agent Roles, Prompt Engineering

LLM Platform

Large Language Models (LLM), Generative AI, LLM Gateway/Routing, Provider Fallback, Structured Outputs, Batch Inference, Guardrails

Retrieval & Knowledge

Retrieval-Augmented Generation (RAG), GraphRAG, Knowledge Graphs, ArangoDB, Vector Databases, Hybrid BM25 + Vector Search

Verification & Compliance

Formal Verification, Lean 4, NIST 800-53/171, MITRE ATT&CK, AI Governance

Document AI & Interfaces

PDF Extraction, Layout & Table Extraction, Extraction Calibration, React, TypeScript, D3, Design Systems

Briefing & Communication

Conference Speaking, Executive & Technical Briefings, Technical Writing

ML & Platform

Machine Learning, Model Fine-Tuning, Classifier Training, Python, Rust, Docker, Linux

DEEPER DETAIL

Omitted from the two-page PDF; kept here for anyone who wants to dig.

• tau — receipt-gated multi-agent harness: agent work compiles to typed DAG contracts with state tracking; every handoff needs a schema-valid receipt or validator result before the next step runs. No receipt, no action.
• ArangoDB agent-memory platform (private, regulated): ~219K evidence-grounded question–reasoning–answer records across 7K+ security-control records (NIST 800-53/171, CWE, MITRE ATT&CK, D3FEND, SPARTA), hybrid BM25 + vector + graph recall over 2.2M source chunks, and ~94K Lean 4 theorem statements indexed for requirements-to-proof retrieval.
• Much of my current client work is export-controlled, so client and program details are withheld; the publicly releasable engineering is at github.com/grahama1970.
• Live capability inventory — skills, sanity coverage, bounded agents, and the research-area map — is generated from the repository at each deploy and published on grahama.co, so the numbers on this page are checkable rather than asserted.

The generated discipline census is technical-inspector material, not resume reading. Inspect technical capability evidence.