Loadwise

Freight brokers were losing loads in their own inbox. A mid-size brokerage came to us with a familiar bottleneck: reps were spending most of their day manually reading quote requests out of email, pulling rates from three different systems, retyping the same load details into a TMS, and calling carriers to confirm appointments. Every extra minute a rep spent on data entry was a minute a competing broker could use to win the load first.

The existing "automation" was a patchwork of email filters and a shared spreadsheet. Nothing talked to anything else, and every new integration meant another manual export/import step. The business wanted an agent that could actually act — read a quote request, price it, and get it in front of a rep to approve — without becoming a black box nobody trusted.

One pipeline, from raw email to booked load. Loadwise is a full-stack platform that watches a shared inbox, classifies incoming messages, and routes them through purpose-built agent workflows: quoting, load building, appointment scheduling, and carrier check-calls. I owned the system end to end — ingestion, backend services, the agent orchestration layer, and the React dashboard reps use to review and approve agent output.

Incoming emails are parsed for shipment details (origin, destination, weight, equipment type, dates), matched against historical rate data and live market pricing, and turned into a draft quote. Reps see the agent's reasoning and sources in a side panel before anything goes out — the agent drafts, the human approves, and only confirmed actions touch the TMS.

Draft everything, commit nothing without a human. The core design principle was separating what the agent proposes from what the system executes. An extraction service turns unstructured email into structured shipment data. A pricing service scores that data against rate history and live lanes. Only after a rep clicks approve does an execution service write to the TMS and notify the carrier — every write is logged with the extracted data, the rate basis, and the approving rep, so nothing enters the system of record without an audit trail.

This kept trust high during rollout: reps could see exactly why the agent priced a lane the way it did, correct it in one click, and the correction fed back into the next extraction pass. Nothing was ever "automated away" without a visible seam a human could grab.

Built for a team that had been burned by tools before. Because the brokerage had tried automation before and abandoned it, adoption mattered as much as accuracy. I built the dashboard to fail loudly and locally: if extraction confidence was low on a field, that field was flagged inline instead of silently guessing. Rate-limited retries, idempotent writes to the TMS, and a fallback manual-quote path meant a bad parse or a downstream outage degraded gracefully instead of blocking a rep's whole queue.

Within the first quarter, the team's own reps were the ones pushing to expand the agent into more of their workflow — the best signal that the tool was actually helping rather than adding a layer to distrust.

Reps went from reading and re-keying every quote request to reviewing drafts the agent had already priced — and within the first quarter they were the ones pushing to expand it into more of their workflow.