Integrations don't fail loudly. They go quiet. Every feed we ship heartbeats in, and silence is the signal. When something breaks, an AI monitor diagnoses it, drafts the exact fix, and emails a GO / NO-GO button.
Nothing runs without a human clicking GO.
Live since August 2026 across a fleet of Windows tasks, Raspberry Pis, cloud services, and databases. The AI proposes; a person disposes.
Scans every 30 minutes. On a failure, an LLM diagnoses it against a measured 117-row map of every system and consumer, drafts the concrete fix — SQL or runbook steps — and emails a GO / NO-GO link. Risk-tiered, with a code backstop so the model can't talk itself past the gate.
Every reachable system pushes success beats; anything silent past twice its interval becomes a diagnosis target automatically. Sources self-register. A dead process or scheduler alarms without having to say so.
A containerized headless agent that diagnoses an engine failure against the repo and runbooks and stages a fix — read-only tools, disarmed output, 900-second and 40-turn caps, a file-based kill switch. Acceptance-tested before it was trusted.
A nightly LLM review of a debt register plus live probes proposes at most one fix per night through the same approval gate — at most three new items, deduplicated. Debt gets paid down instead of catalogued.
Purpose-built detectors for the case a failure alert can't catch: the job that ran, exited clean, and did nothing. Print spools, invoice queues, import chains — with self-healing task restarts within ten minutes of a death.
Health state for every integration source — storefront, marketplace, payments, mirror freshness, shipping, documents — feeding the advisor's scan every cycle, so the diagnosis starts from the whole picture.
Gmail, Microsoft 365, IMAP, Google Forms, and the copier — every place paper and email enter a business, turned into structured data and routed by rule.
Pallet photos and prep sheets from a warehouse tablet form OCR'd and routed by a five-rule priority — broker, portal, or collect — with a portal-exclusion list and an approval loop.
Multi-page copier scans classified page by page; order, BOL, and customer extracted into a searchable index. One rebuild lifted fully-keyed packets from 44% to 81% and PO accuracy from 20% to 95%.
PO receipt scans matched to open purchase orders; outgoing PO PDFs parsed to keep open-PO data current — on the same 30-minute weekday cron fleet, no re-keying.
Per-invoice email time cut from ~15 seconds to ~1 with a queue and single-session flush; a one-process dispatcher drains a batch in dependency order — 43 invoices in 82 seconds where the old engine took 14 minutes for 44. Proven by delivery, not logs.
The ERP's invoice and pick-ticket documents regenerated byte for byte from data, regression-tested against a 128-stream gold corpus. Rules derived from real edge cases, not the manual.
Carrier invoices ingested, OCR-extracted, and routed through an approval queue to payment — a queue, detail, and bulk-approval UI over a real backend, wired to the payables platform and the accounting system.
Approval-first, everywhere. Even the mailbox scout that proposes calendar entries for facility visits keeps a human reply as the judgment layer — no LLM in the decision path. Automation earns trust by asking.
Every integration we build ships with monitoring from minute one. If you already have integrations and no idea whether they ran last night, that's where we start.