Engineering methodology for deploying AI systems with operational discipline.
A system that ships predictably beats one that demos brilliantly.
Most AI projects ship as demos and die in production. The failure modes are boringly consistent: costs spiral because no one tracks token spend, pipelines break halfway through and restart from scratch wasting compute, and multi-agent systems bottleneck on a single supervisor that burns tokens just watching. These aren't capability problems — they're operational problems.
The Human Stack is my response: a living engineering methodology that treats AI systems as operational systems. Every claim is evidence-graded (E0–E5). Every architectural choice is append-only, dated, and cross-referenced in an ADL. Every tool carries adversarial tests before it ships. The methodology is public in the-human-stack.
The Arete Stack is its implementation — the tools I actually build and maintain. Each repo maps to a layer of the methodology: skills for session discipline, cost observability for spend anomalies, governance linting for workflow anti-patterns, orchestration for autonomous improvement, and the methodology reference itself. This profile is arranged as an onion — start at the outer layer and peel inward.
Before software I spent 17 years in manufacturing and logistics (IBM, Toyota Production System). That discipline shows up in every codebase: standard work, visual management, error-proofing, continuous improvement.
Substack · LinkedIn · jamesyng79@gmail.com
- The Eval Is the Product — Why evaluation infrastructure matters more than model choice.
- Local-First AI Stack — Zero-cost inference with Ollama on consumer hardware.
- Decision Logging as Operational Memory — The ADL methodology: append-only, evidence-graded, revisit-conditioned.
These are the repos I maintain to production quality — evidence, tests, and documentation first.
| Project | Stars | Language | What It Is |
|---|---|---|---|
| animus | ⭐1 | Python | Personal AI operating environment — bitemporal memory, evidence-graded maturity, autonomous improvement |
| mcp-manager | ⭐1 | Python | MCP server lifecycle manager across Claude Code, Cursor, Windsurf — now with permission-prompt audit |
| ai-spend | ⭐0 | Python | Cross-provider AI cost aggregation with provider retry, config encryption, and health checks |
| agent-lint | ⭐2 | Python | Workflow YAML cost estimator + anti-pattern detection for agent orchestration |
| the-human-stack | ⭐0 | Markdown | Living engineering reference — evidence-graded chapters, benchmark methodology, failure taxonomy |
| ai-skills | ⭐3 | Python | Production-ready skills for Claude Code personas and orchestrated workflows |
| arete-evals | ⭐1 | Python | LLM eval suites and run records — eval engineering vs eval practice |
Hobby projects (EVE Online lore, personal context servers) are kept private. What you see here is what I ship.
The fastest way to understand the methodology is to use it. These repos are arranged as an onion — start at the outer layer and peel inward.
| Layer | Repo | What It Is | Entry Point |
|---|---|---|---|
| Skills | ai-skills | Production-ready skills for Claude Code, agent orchestration, and workflows | ./tools/install.sh --bundle arete-studio-ops |
| Eval | arete-evals | LLM eval suites and run records — eval engineering vs eval practice | pip install -e . |
| Cost | ai-spend | htop for AI spend — cross-provider cost aggregation |
pip install ai-spend |
| Governance | agent-lint | Workflow YAML cost estimator + anti-pattern linter | pip install agentlinter |
| Orchestration | animus | Personal AI operating environment with local-first control, evidence-graded maturity, session persistence, autonomous improvement | git clone + docker compose up |
| Methodology | the-human-stack | The engineering reference itself — evidence-graded chapters, benchmark methodology, failure taxonomy | Read VISION.md |
Animus is the primary reference implementation — a personal AI operating environment with local-first control and evidence-graded maturity tracking.
| Subsystem | What It Does | Tests |
|---|---|---|
| Kernel | Bitemporal memory core, checkpoint/resume, session persistence | 179 green |
| Head | Quality gates, fallback routing, natural-language interface | 72 green |
| Citizens | Architect (autonomous improvement proposals), Conversation Designer, Knowledge Curator | 35 green |
| Forge | Eval pipeline, benchmark execution, rubric-based scoring | Integrated |
| Session Controller | Token-budgeted session lifecycle, graceful wrap, auto-restart | 22 green |
Key design decisions (visible in the code):
- Evidence Framework — Six-stage maturity model (Concept → Self-improving) with Coverage KPI. Makes "documented but unverified" into "inspectable evidence."
- ADL-governed architecture — Every major decision is append-only, dated, and cross-referenced. No tribal knowledge.
- Quality gates before merge — Deterministic scoring (tool/completeness/structure) with adversarial test harness. Features don't ship without gate passage.
| Tool | What It Does | Install |
|---|---|---|
| ai-spend | htop for AI spend — cross-provider cost aggregation with OpenRouter, Anthropic, OpenAI |
pip install ai-spend |
| mcp-manager | MCP server manager across agentic IDEs (Claude Code, Cursor, Windsurf) | pip install arete-mcp |
| agent-lint | Workflow YAML cost estimator + anti-pattern linter | pip install agentlinter |
| context-hygiene | Heuristic context-window analyzer — position-decay scoring + regex patterns | pip install context-hygiene |
| arete-evals | LLM eval suites with benchmark scoring, A/B comparison with bootstrap CIs | pip install -e . |
| ci-templates | Reusable GitHub Actions workflows, lint configs, release automation | Copy + adapt |
| RedOPS | Red-team playbook for AI infrastructure — incident response, supply-chain audit | Read + run |
| overwatch | Multi-service health monitoring with structured logging and alerting | pip install -e . |
- Decision logs, not memory. Every architectural choice is an ADL entry — rationale, rejected options, tradeoffs, kill criteria.
- Tests are deliverables. Not an afterthought. Adversarial tests prevent regression of quality-gate contracts.
- Local-first where possible. RX 7900 XTX inference stack (Ollama, 4-model tiered) for zero-cost eval calibration.
- Ship, measure, error-proof, repeat. Kaizen learned from scaling an ice-cream line 6× — applied to software.
Direct entry points for "what does the code actually look like":
- Animus Kernel — bitemporal memory + adversarial tests — v2.3 scaffold: valid-time / transaction-time axes, quality-gate contracts enforced before any feature ships. Architect Citizen produces ImprovementProposals from codebase observation.
- ai-spend — OpenRouter provider —
GET /api/v1/generationspagination withnative_costfallback to token-based estimates. Pattern: provider registry ABC with side-effect registration. - Animus Session Controller — Token-budgeted session lifecycle (96% utilization trigger), graceful finalization with model-generated summary, checkpoint + auto-restart. 22 tests, 83 existing head core tests green.
- Animus P5 Discovery — MCP scanner, OpenAPI ingestion, annotated script discovery, 4-dimension schema validator, hash deduplication. 41 tests, 194 total P0–P5 tests green.
17 years in manufacturing and logistics operations (IBM, Toyota Production System). Scaled an ice-cream production line from 740 pints/day to 4,800/hour using Kaizen. Now apply the same discipline to AI infrastructure: standard work, visual management, error-proofing, continuous improvement.
See it through. Do it better. Leave something real.