Automation that survives production.

Invoices and purchase orders arrive by email; validated structured data leaves in a database. Every extraction is checked mechanically — do the line items sum to the total, is the date parseable, is the vendor known — and anything below the confidence threshold routes to a human review queue with the original document side-by-side.

• Validation is arithmetic, not another LLM call — totals must reconcile before a record is written.
• Idempotent writes: re-running a batch can never create duplicate records.
• Extraction accuracy is tracked weekly against the review queue's corrections.

Answers the phone, qualifies the caller, checks live calendar availability, books the slot, and sends an SMS confirmation. Built for businesses that lose money on every missed call. The agent's tools are narrow by design — it can check availability and book, nothing else.

• Sub-800ms response latency budget — beyond that, callers hang up.
• Hard escalation rules: two failed intents or any complaint transfers to a human line.
• Every call logged with transcript, outcome, and per-call cost.

The layer that turns workflows into a service: self-hosted n8n in queue mode with worker replicas, a global error workflow that catches every failed execution, retries with backoff, and dead-letters what can't recover. Per-client credential isolation, per-client cost tracking, and a one-page dashboard the client can actually read.

• Workflows are versioned in git and deployed, not edited live in production.
• Failures are loud: every dead-lettered execution alerts Slack with full context.
• Monthly report per client: executions run, hours saved, errors, spend.