THE PROBLEMEquipment repair shops lose critical repair days when parts run short against open jobs, and owners cannot trust a black-box AI with the shop checkbook—one stale stock count, unapproved broker, slow lead time, or hallucinated PO can blow a deadline or overspend thousands.
OUR SOLUTION — "THE MODEL NEVER HOLDS THE CHECKBOOK"
Parts Buyer is the shop owner's personal purchasing agent. In one click, it checks inventory (P01–P08) against open repair jobs (J1–J4), evaluates offers by true landed cost (unit price × qty + shipping, MOQs, bulk tiers), rechecks live supplier stock/price before buying, auto-places orders within the spend limit, pauses over-limit orders for 1-click approval, and writes a plain-English audit trail ("receipts"). Rules R1–R7 run deterministically in rules.py while daily idempotency keys (YYYY-MM-DD:PartID) guarantee zero duplicate orders across retries or Crash test runs.
SPONSOR TOOLS USED
1. Amplitude + Statsig (config_flags.py): Live spend_limit read from Statsig Dynamic Config parts_buyer_rules via statsig-python-core (polled every 2s)—raising it from $1,000 to $3,000 flips P06 ($2,395) from pending_approval to placed live with no redeploy.
2. Cognee Cloud (memory.py): Ingests docs/policy.md and logs run-over-run supplier lessons (/api/v1/remember, /recall, /search) into the audit trail + Inspect Cognee modal.
3. Firedrill (simulations.py, firedrill.recipe.json): 7 stress-test drills scoring R3, R5, R6, R7—7/7 passed (100%), 0 violations.