APilot

Deterministic decision core (Python): three-way matcher (apilot/matcher.py) with ±0.5% price tolerance, duplicate-invoice and uniform ~10% tax-uplift detection, per-finding severity; decision engine (apilot/decide.py): no findings ⇒ AUTO_POST (confidence 1.0), any finding ⇒ HUMAN_REVIEW (confidence 0.0) with a suggested resolution.
Audit & evaluation: every decision written to an audit trail (data/audit.json); a read-only evaluator scores the trail against ground-truth labels with per-type precision/recall/F1.
API (FastAPI): summary, invoice queue, review actions, capability and evaluation endpoints; also serves a single-file HTML dashboard.
Control Desk UI (Next.js 15 / React 19 / TypeScript / Tailwind / Radix): four metric cards (processed / auto-posted / unresolved / reviewed), a searchable queue with Unresolved / Reviewed / Auto-posted scopes and a failed-policy filter, an invoice detail view with live posting status, failed policy, finance owner, evidence (actual vs expected), and a recommended action, plus approve-exception / hold-payment / escalate decisions that require a recorded reason. An optional extraction card (dry run) is disabled until the backend reports it enabled.
Built with Agent Orchestrator: an orchestrator plus parallel worker sessions, each worker in its own isolated git worktree and branch, test-gated before landing (68 pytest tests passing; live session/commit counts on the AO dashboard / git history).