Production-minded reference workflow for AI coding and support-ticket automation. It is intentionally small, but it demonstrates the engineering surface employers usually look for after the first headline: typed contracts, local RAG, safety gates, confidence scoring, structured output, evaluation, traceability, CI, and documentation.

This is not a chatbot demo. The workflow is deterministic so it can be tested without live model access, then swapped behind explicit agent/model interfaces in a production deployment.

• src/ Python package layout with Pydantic contracts
• local Markdown knowledge base as the only authoritative source
• retrieval with exact error-code boosting
• classification, clarification, refusal, and resolution states
• prompt-injection and unsafe-request detection
• structured TicketResolution output
• confidence calculation based on evidence, metadata, validation, and safety
• JSONL OpenTelemetry-style trace events
• regression eval cases and CI quality gates

SupportTicket
  -> intake and safety screen
  -> local KB retrieval
  -> deterministic classifier
  -> confidence and policy gate
  -> schema-validated TicketResolution
  -> JSONL trace events

The same seams map cleanly to LLM-backed agents later: prompts live in prompts/, workflow state is explicit, and final outputs are validated before any downstream system consumes them.

uv sync --locked --extra dev
uv run python scripts/run_demo.py
uv run python scripts/run_evals.py

Example output is a structured ticket resolution with category, priority, status, proposed solution or follow-up questions, retrieved sources, safety flags, and reasoning trace.

uv run ruff check .
uv run ruff format --check .
uv run mypy src
uv run pytest
uv run coverage run -m pytest
uv run coverage report
uv run python scripts/run_evals.py

The GitHub Actions workflow runs the same gates on push and pull request.

This repository is a compact example of the same production AI patterns used in larger systems:

• governance: AgentGate
• incident agent orchestration: Incident Command Mesh
• retrieval evaluation: Metivta Eval

• src/ai_coder_production_lab/: workflow, contracts, retrieval, safety, tracing, evals
• knowledge_base/: local authoritative support knowledge
• prompts/: versioned prompt assets and lifecycle notes
• tests/: unit, workflow, tracing, and regression-eval tests
• docs/: architecture, RAG, observability, threat model, and trade-offs
• schemas/v1/: versioned response contract for downstream clients
• .github/workflows/ci.yml: lint, format, type, test, coverage, and eval gates

Production AI coding systems fail less because of model choice than because of weak boundaries: untyped outputs, hidden state, ad hoc prompts, no evals, no policy gate, and no traceability. This lab keeps those concerns visible in a runnable codebase.