Seven years moving freight across the MX-US border. Now shipping AI agents that run a real operation.

Problem

Cold outreach for a freight operation cannot be "an AI that writes emails." It needs research that is actually true, copy that does not read like AI, and hard limits so an autonomous system never damages the sender's domain or reputation.

What I built

A multi-agent pipeline that runs the entire outbound motion. A researcher agent enriches prospects and verifies a concrete, sourced hook for each one. A writer agent drafts a 4-touch sequence in the sender's voice. Two adversarial gates review every batch: a technical reviewer hunting fabricated claims, cloned paragraphs and sequence violations, and a "customer" agent that role-plays the actual recipient (a supply chain director who already has forwarders) and votes REPLY or DELETE. An operator agent sends.

The part that matters: limits live in code, not in prompts. The send script mechanically enforces timezone-anchored sending windows, daily ramp caps, a global suppression list and hard-blocked domains. State is a JSON ledger that only trusts the mailbox as ground truth. Anti-clone checks are scripted: no long phrase may repeat across a batch, and every relative time marker ("last year") is validated against the source's date.

Results

Full 4-touch sequences delivered with zero bounces and zero unsubscribes. When a bounce appeared that no agent had logged, the attribution check surfaced the unregistered send path within a day. An earlier incident on a shared mailbox became a coordination layer of its own (case study 04).

Problem

In logistics, response speed wins deals: a shipment stuck on a Saturday cannot wait until Monday. But full autonomy is dangerous. An agent that invents a freight rate destroys trust instantly.

What I built

A closer agent that reads every inbound reply and web lead for the freight operation, classifies the actual intent of the thread (not keywords), and decides against an explicit autonomy matrix. Green cases and yellow cases without numbers it answers and sends on its own. Anything involving pricing or risk it drafts, escalates and waits.

One rule is enforced in code: the agent can never state a rate. It collects shipment details (lanes, weights, dimensions, frequency) and escalates the number. Every thread it works becomes an append-only lesson log, so the next reply closes better than the last.

Results

Inbound replies get worked to a meeting or a quote request without a human in the loop for the safe cases, at any hour, which in cross-border freight is a genuine differentiator. Zero invented rates and zero red-category messages sent autonomously, by construction.

Problem

At a freight operation, carrier payments arrive by email: a PDF remittance with loads and amounts that someone must reconcile against the TMS and register, load by load, without paying twice, without paying the wrong amount, and without missing the ones that land late on a Friday night.

What I built

A back-office agent that runs unattended on a schedule, with several passes on payment days. It reads the payment emails, parses the remittance PDFs, reconciles each load against the TMS via API, registers the payment and uploads the receipt. A deterministic script handles everything routine. An LLM "brain" is invoked only for the cases that need judgment: a batch the parser rejected, a bank amount that does not match the TMS, a load with two receipts.

The principle that makes real-money autonomy safe: the brain decides which case applies; a separate script decides if it is allowed. The apply layer enforces code-level locks: payment-date cutoffs, the receipt PDF as the only valid date source, the TMS as the only valid amount source (never the email, never arithmetic), mandatory reason strings for any explained discrepancy, already-paid detection, sender scoping, and a hard rejection for anything else. An explicit autonomy matrix splits cases into green (apply alone), yellow (apply and notify in the same pass) and red (write a proposal, notify, stop), red covering possible duplicate payments or anything requiring a reversal. Every run appends to a log with signals a human can scan in seconds.

Results

Unattended production operation over real carrier payments since June 2026, with zero erroneous payments. The worst failure mode observed to date is an expired session flagged in the log for a human to fix. The system's autonomy was widened once, deliberately, after a stretch of clean operation.

Problem

Multiple independent AI sessions (prospecting sequences, a campaign engine, one-off kits) all sent email from the same mailbox and could not see each other. One day the mailbox closed at 83 sends against a ramp limit of 30, and a follow-up wave landed on a lead who had already replied.

What I built

A coordination layer every sender must call before sending. estado reports today's true count and who sent what. reservar claims a quota slot per session. filtrar joins three state sources nobody had crossed before (global suppression, sequence state for replied, bounced and active leads, and today's sends) and removes anyone who must not be emailed. cerrar reconciles reservations against reality.

Counting runs against the mailbox as ground truth, by calendar day in the operation's timezone, initiative emails only (replies do not consume ramp). That replaced a naive rolling 24-hour query that proved wrong twice, over-counting by 26%. Any send the ledger cannot attribute is flagged as a rogue path to hunt down.

Results

The over-send never recurred. The "unattributed send" alarm caught a real unregistered send path within a day of shipping. Incident, root cause, mechanism: all documented.

The background

Before building AI systems, I operated and sold the thing they automate.

• Seven years in international logistics: cross-border Mexico / US / Canada through the Laredo and Colombia crossings, plus a domestic Mexico operation at 190 shipments per month.
• FTL and specialized freight, including data center equipment: the physical infrastructure the AI industry runs on, moved across a border.
• Sales side: 17 enterprise accounts closed, $72M+ USD revenue over 5 years, largest single account at $1M USD per month.
• Industry specialization in electronics and tech manufacturing, the same vertical enterprise AI deployments target.
• Fluent in the operator's stack (TMS, tracking, pricing, ERP) and in both languages of the border, on both sides of the desk: operations and sales.