brain/docs/adr/ADR-007-ml-ai-tooling.md

ADR-007: ML / AI tooling (A11)

• Status: Accepted
• Date: 2026-05-17
• Deciders: Дмитрий

Context

The A11 «ML / AI-разработка» map section had zero tooling. Лидерра ships no ML/AI code; calc_lead_score is a deterministic SQL function. A toolset is needed for the day AI features (LLM-backed) or a scoring model are scoped.

Decision

A11 adopts a six-position toolset in two subcategories:

• LLM integration — the claude-api skill (build), promptfoo (test prompts), Sentry MCP (observe). All reuse or new-light.
• Classical ML — a vendored Data Scientist skill (workflow knowledge). The executable part, Jupyter MCP, is deferred: it needs a Python ML runtime the deliberately-minimal native-Windows machine lacks, and there is no model to train. Jupyter MCP is a reserved registry slot, installed by a separate task when a concrete model is scoped.
• promptfoo runs manually / CI only — never in a hook (paid LLM calls).
• A11 tools are non-UI → the ml-ai-tooling off-phase category.

Consequences

• Positive: A11 populated; AI features have a build+test+observe toolchain.
• Risk: the Data Scientist skill is third-party (CC BY 4.0 content) — mitigated by vendoring with attribution into .claude/skills/data-scientist/.
• Cost: promptfoo is a heavy devDependency (~1090 transitive packages, one native module). Accepted — it is dev-only tooling, not shipped to the app.
• Deferred: no Python runtime until a model is scoped — accepted, this is the decision.

Enforcement

None — A11 tools are advisory; verified by use and code review.