AI Architect - Native Integration

The layer where every carrier, port, forwarder, and customer ERP disagrees. Ingest against real-world hostility, reconciliation when three sources disagree, freshness as a stated commitment, and a write path that cannot be half-complete. Plus the question that makes this different from a classic integration role: what a model is allowed to do across that boundary, and what happens when it is wrong. It is not the Architect/Senior Engineer role; two roles here carry the word "architect, " and they meet at exactly one place, the contract an agent calls. They own how our own system is decomposed; you own the outside edge.

The work

Responsibilities:

• Build ingest for EDI and flat files alongside REST, with no versioning discipline on the other side: silent schema drift, duplicate and out-of-order events, and 200-with-garbage as a normal response. Detect it rather than trusting it.
• Design reconciliation for the same shipment, three sources, and three different arrival estimates. Which wins, how the decision is recorded, and what the customer sees while it is unresolved.
• Own freshness as a number: a staleness threshold per source, and an alert on the absence of data rather than only on errors.
• Use models where they earn it: mapping a new partner's fields, classifying a document, extracting from a PDF with typed output, a validation gate, and a review path for low confidence. Then state the accepted error rate and who absorbs the errors.
• Make agent-initiated writes safe: idempotency keys that reach the third-party call, deduplication, replay, and reconciliation after a partial failure. Booking a truck twice costs a real person real money.

Requirements:

• Five years is a floor. Integration scar tissue counts for more here than title history if you have it; apply.
• 5+ years building integration or platform infrastructure in production, including an API consumed by callers you did not control.
• Exactly-once-effect thinking in practice: idempotency keys, deduplication, replay, dead-letter handling, and reconciliation after a partial failure.
• Credential depth: OAuth and token lifecycles, rotation, tenant-scoped permissions, least privilege enforced per integration rather than per platform.
• Event-driven infrastructure queues, change data capture, webhooks with an explicit position on ordering, late arrival and backfill.
• Experience with counterparties that are slow, inconsistent, or undocumented. Carrier APIs, EDI, port and customs systems, or ERPs are directly relevant.
• You can design an API for a caller that never tires, never reads prose, and cannot be trusted to notice a half-completed write.
• You will be on call for the integration layer and lead its incident reviews.

Also good, none of it required:

• You have designed a permission model that had to survive an enterprise security review.
• Logistics, trade, payments, or another domain where a duplicate write costs real money.