AI Enablement / Forward-Deployed Engineer — Portland, OR

I ship LLM-powered products end to end — eval harnesses, multi-agent systems, billing, deploy — and I have the operations background to make them reliable. Before software I spent 17 years running manufacturing and logistics systems (IBM, Toyota Production System): standard work, visual management, error-proofing. A system that ships predictably beats one that demos brilliantly.

Portfolio · LinkedIn · jamesyng79@gmail.com

Python · TypeScript · Rust · React / Next.js · FastAPI · PostgreSQL / SQLite · LLM evaluation & routing · multi-agent orchestration · MCP servers · Stripe · CI/CD (GitHub Actions) · Vercel / Fly.io · Azure

20+ public repos · 15+ published PyPI packages · 15,000+ tests across public packages · 5 MCP servers shipped · live Stripe billing

What I'm building right now (Derek Sivers style):

• The Human Stack — A living engineering reference for deploying, operating, and evaluating AI systems. Evidence-graded chapters, engineering reviews, benchmark methodology, and deployment case studies. The intellectual center of my public work.
• Animus v2.3 — Sovereign AI operating environment and primary reference implementation for The Human Stack: multi-agent orchestration with budget controls, quality gates, checkpoint/resume, and autonomous self-improvement. 55K+ LOC, Evidence Framework maturity tracking.
• Crucible — Phase-gated transformation framework: four/five conditions, structured failure taxonomy, active/receptive polarity. Research layer under the Animus Mind-class scaffold.
• Local AI Stack — RX 7900 XTX powering zero-cost inference (Ollama, 4-model tiered stack). Eval calibration runs weekly against arete-evals suites.
• TIAID consulting — $2,500 Rapid Assessment + $15K–$25K Full Engagement products for trauma-informed AI deployment inside organizations.

If you only look at three things, use this path:

1. The Human Stack — My living engineering manual: evidence-graded methodology for deploying AI as operational infrastructure. Start with the Engineering Reviews and Case Studies.
2. Animus — The sovereign AI operating environment that implements the principles in The Human Stack: multi-agent orchestration, autonomous self-improvement, and integrated governance.
3. ai-session-templates — Structured session templates for Claude Code, Codex, and repo-aware coding agents.

If one helps you, please star it. If something breaks, open a setup-blocker issue and I will prioritize it.

BenchGoblins — AI fantasy-sports decision engine. Scored LLM routing under the hood, commissioner tools, live on Fly.io + Vercel with Stripe billing. Production codebase private; comparable scored-routing patterns visible in memboot.

EVE Gatekeeper — Route-intelligence platform for EVE Online: interactive 14-layer map, per-hop risk breakdown, gate-camp warnings. Stripe billing live. GitHub

Past: anchormd — AI-agent context-file generator and auditor (CLAUDE.md / AGENTS.md). Archived 2026; lessons absorbed into Animus document-control system and the ai-session-templates builder pipeline.

The Human Stack — A living engineering reference for deploying, operating, and evaluating AI systems. Evidence-graded chapters (E0–E5), engineering reviews with retrospectives, benchmark methodology, and deployment case studies. The intellectual center of my public work; everything else is evidence supporting it.

Animus — Sovereign AI operating environment and primary reference implementation for The Human Stack: multi-agent orchestration with budget controls, quality gates, checkpoint/resume, autonomous self-improvement, and Evidence Framework maturity tracking. ~17K LOC across core packages. Previously private; now public and actively developed.

Animus Mind — v2.3 Mind-class architecture: bitemporal memory core, adversarial tests, deterministic quality scoring, Architect Citizen for autonomous improvement proposals. The reasoning layer behind Animus.

memboot — Zero-infrastructure persistent memory layer for any LLM — works with Claude Code, OpenAI Codex, Cursor, Windsurf, Claude Desktop, and Zed. Semantic-security audit caught two MEDIUM findings the SAST run missed.

arete-evals — Public eval-suite records and run artifacts from the Animus Forge calibration pipeline. Bootstrap A/B comparison, weekly calibration, rubric-based scoring.

RedOPS — Professional cybersecurity intelligence & attack surface management platform. OSINT automation, MITRE ATT&CK threat-path mapping, risk quantification (likelihood × impact), and executive-ready reporting. Strict scope enforcement and audit trails for authorized defensive assessments.

overwatch — Tactical ISR dashboard — unifies YOLO object detections, OSINT intel feeds, and drone telemetry into a single operational picture. Entity resolution, auto-briefing SITREP generation, geofencing, real-time WebSocket feed, and an 8-tab Streamlit dashboard.

chainlog — Tamper-proof audit trails for AI agents on Base L2 (Ethereum). Writes cryptographic fingerprints of actions on-chain — no PII, just hashes. Includes TypeScript + Python SDKs, CLI verifier, and a Next.js dashboard. Model version pinning for EU AI Act compliance, dead man's switch for contingency triggers.

stellar-audit-agent — Pay-per-call AI code audit API with dual payment rails: x402 micropayments + Stripe MPP on Stellar. Autonomous Claude-powered agent discovers services, fetches repos, reasons about audit scope, pays per-request, and synthesizes results. Live demo on Fly.io. Launch demo

I ship MCP servers that give AI assistants operational superpowers:

Aurora Arcology — Investigation-board framework for narrative universes: an interactive corkboard of nodes, sourced claims, and confidence-weighted connections. Next.js 15 + TypeScript + SQLite/Drizzle, runtime-editable ontology. GitHub · Live demo

ai-skills — Production-ready skills for Claude Code and multi-agent systems.

ai-session-templates — Structured session templates for Claude Code, Codex, and repo-aware coding agents.

Argus Overview — Linux multi-window manager for EVE Online. PyPI · 26K+ downloads.

Dossier — Local-first document intelligence: ingest PDFs/emails/scans, extract entities, surface relationships, forensics timeline. Patterns extracted into Animus Mind v2.3 (entity resolution, provenance tracking, forensics timeline).

EVE Frontier tooling — Monolith: on-chain anomaly detector for EVE Frontier on Sui, with a live 3D-map demo of 24k systems.

I treat test coverage and eval calibration as first-class deliverables. Here's the public record:

Latest calibration run: 2026-07-02 — config_loader + rate_limiter test cases under repair; weekly schedule resumes after fixes. View history →

Direct entry points for "what does the code actually look like":

• memboot v0.7.1 — SSRF guard + credential-dir denylist — Semantic-security audit caught two MEDIUM findings the SAST run missed. Shipped scheme allowlist with redirect re-validation + extended default ignore_patterns to skip credential directories. Regression test asserts the exact attack scenarios stay out of discover_files output — poka-yoke against silent regression of the default skip list.

• Argus Overview — character-logoff detection — Spec-driven feature in a 26K-download tool: tracker subscribes to existing character_gone signal, idempotent slot, 11 new tests including p95 latency under 5ms across 100 trials. Architecture luck — the detection signal already existed; the work was wiring + verifying.

• aurora-arcology — Dossier integration scoping — Cross-project leverage analysis: four bridges from Dossier (forensic NER + briefing endpoint) into Aurora (narrative-investigation board), ranked by ROI with dependencies + effort estimates per bridge. The FDE pattern of recognizing where one product's primitives serve another's gap.

• Animus Mind — bitemporal core + adversarial tests — v2.3 scaffold: bitemporal memory model with valid-time / transaction-time axes, adversarial test harness asserting quality-gate contracts before any feature ships. Architect Citizen produces ImprovementProposals from codebase observation.

• The Human Stack — Evidence Framework & Benchmark Methodology — Reproducible evaluation across four dimensions (correctness, precision, efficiency, evidence quality), with rubric-based scoring, failure taxonomy, and continuous drift detection.
• chainlog — On-chain vs Off-chain Audit Design — Only hashes go on-chain (Base L2); zero PII leakage. Model version registry for EU AI Act compliance, dead man's switch for contingency triggers.
• RedOPS — MITRE ATT&CK Mapping & Risk Scoring — Structured threat-path generation with constrained output validation. Likelihood × impact quantification with audit-scoped enforcement.
• Animus Mind — Bitemporal Memory Model — Valid-time / transaction-time axes adapted from temporal database research (Snodgrass & Jensen). Adversarial test patterns prevent regression of quality-gate contracts.
• Animus Evidence Framework — Six-stage maturity model (Concept → Self-improving) with Coverage KPI. Makes "documented but unverified" into "inspectable evidence."

TIAID — Trauma-Informed AI Deployment — a methodology for rolling out AI inside organizations without breaking the people, mapped to the NIST AI Risk Management Framework.

I scaled an ice-cream production line from 740 pints/day to 4,800/hour using Kaizen — and I bring the same discipline to software: ship, measure, error-proof, repeat. See it through. Do it better. Leave something real.