ROED
RFQ email in, priced quote out. Claude reads the request, the forwarder's own rate card prices it, and a person approves every quote before it leaves. A rule handles it, or a person does. Never a guess.
A shipper emails a request for quote. The forwarder opens Estes, then Old Dominion, then FedEx Freight, then does the math in a spreadsheet. That is forty minutes for one quote, and it happens twenty, thirty, forty times a day. Often the shipper never replies. The forwarders doing this are mid-sized, $10M to $200M a year, running on a shared inbox and spreadsheets. The freight AI platforms are built for companies ten times their size, so this segment gets nothing.
Split the job by what each part is good at. Claude reads the email, because reading messy human writing is what language models do well. Everything that touches money is plain code: the price comes from the forwarder's own carrier rates, freight classes, zones, fuel surcharge and margin rules, never from the model. And a person signs off on every quote before it goes out. ROED was built with beta testers at a domestic forwarder, whose real quoting workflow and pricing tables shaped the engine.
- Email in, three ways: a Front webhook, a Front poller, or Nylas for Gmail and Outlook; every webhook is signature-checked, and auto-replies and out-of-office messages are filtered before any AI call
- Field extraction: Claude Sonnet reads the email and returns twelve shipment fields (origin, destination, weight, freight class, pickup date and more), each with its own confidence level
- Hazmat cross-check: a second, separate Claude call re-reads any request flagged as hazardous before it can move forward
- Confidence gate: if any of the five required fields isn't high confidence, or the load is hazmat, multi-shipment, full truckload or has a past pickup date, it goes to a person instead of the rate engine
- Deterministic rate engine: freight class and weight bracket, a zone multiplier, accessorials, the weekly EIA diesel surcharge, per-leg margin and a minimum charge, all from the forwarder's own rate card
- Multi-leg quotes: first-mile, line-haul and last-mile bundled into one price when the forwarder has carriers for each leg
- Human approval: the dispatcher reviews the quote and can adjust it by up to 15% either way, with a reason code that lands in the audit log
- In-thread sending: the quote replies on the original email thread from a template; no AI-written text reaches the customer
- Customer accept or decline from a secure link, then a generated Bill of Lading once the pickup leg is booked
Live in early access at getroed.com, designed to take a quote from inbox to sent in under three minutes.
Next.js 16 · React 19 · Tailwind v4 · Vercel
Python FastAPI on Railway
Supabase: Postgres, Auth, Storage; every table scoped by tenant
Claude Sonnet: twelve fields with per-field confidence; temperature 0; prompt caching
Claude Sonnet: a second, independent read of hazardous loads
Deterministic Python rate engine; no model in the pricing path
Front (webhook + poller) · Nylas (Gmail, Outlook)
EIA weekly diesel price, applied as a surcharge
Bill of Lading PDF, with 49 CFR certification wording on hazmat loads
Sentry
Four decisions that don't show up in a feature list. Keep the model away from the money: there are exactly two Claude calls in the whole product, both reading, neither writing. No price, no surcharge and no customer-facing sentence comes from a model, so a hallucination can never reach a shipper's inbox. Make confidence mean something: each field's confidence describes what the email actually says, not how sure the model feels, and the gate turns it into one plain decision, price it or show a person. Learn from every correction: when a dispatcher fixes a field, the confidence it had at that moment is saved, so correction rates can be read per confidence level. It informs the thresholds and never tunes itself. Hunt for silent failures: a drift test fails the build if a pricing path stops running or the database starts rejecting values without an error.
ROED is the clearest case for putting AI exactly where it's strong and nowhere else. The tempting build is one big agent that reads the email, prices the load and writes the reply. It would demo beautifully and fail quietly on the one thing a forwarder can't afford to get wrong. Instead the model reads, the code prices and a person approves. The product's own line says it best: a rule handles it, or a person does. Never a guess. That's the shape Barael builds for any business where a wrong number costs real money.