Kyle Rivelli.

AI automation consulting

I build AI automation that runs real operations.

I automated the back office of a freight brokerage from inside the ops seat: two email bots, four API integrations, three unattended crons, all running on live commercial data. The AI reads. Deterministic code decides. Humans hold the kill switch.

loads posted, zero corrupted
4,342
first ETA cycle: 35 asks, 16 confirmed, 3 real flags
40 min
to parse one carrier email
~$0.003
checks passing before the first live send
223

// every number sourced from production logs, commit history, and pre-flight audit records

Shipped and running

Six systems on live freight

Fourteen production builds in 33 days for a freight brokerage. These are the six that carry the day-to-day.

Why it can be trusted

The method is the product

The AI is boxed in

Claude reads emails and classifies replies. It never writes a message on its own authority, never routes, and never touches a price. If the model is wrong, the blast radius is a mislabeled reading, not a wrong number on the wire.

One-way doors get double guards

Anything irreversible re-reads live state immediately before the write, and writes are read back to confirm they stuck. A timer alone can never fire an action that has no undo.

Shadow mode before trust

Every bot launches writing drafts, not sending. Independent kill switches, all defaulting to off. Going live is a human decision made after watching real traffic, not a config default.

Some of your back office is buildable.

If your team is paying someone to post listings, chase status updates, or re-key one system into another, some of that is buildable. I can usually tell you which part in one conversation.

Or start in writing

Tell me about the manual work