← Platform
AI on the platform

AI built for operators —
not for demos.

Toretto's AI surfaces aren't bolted-on chat widgets. Every agent runs through a permissioned tool layer, writes through human- confirmed workflows, and leaves a per-step reasoning trail you can audit. The model is Claude. The discipline is ours.

Agent Runs list with task type, source document, duration, turns, and token counts

Every agent run is tracked — task type, source, turns, tokens, status. Click into any one to see the reasoning trail.

What the AI does on the platform

Five concrete capabilities shipping today.

Rate interpretation

Read any rate card

Upload a UPS, FedEx, DHL, or LTL tariff in PDF or XLSX. The rate-interpretation agent parses it into structured rate tiers — base rates, zones, weight breaks, accessorials, dimensional- weight rules. You review the agent's interpretation before activating, and reparse anytime the carrier sends an update.

  • Multi-page tariff documents
  • Mixed table + prose formats
  • Per-customer override schedules
Invoice audit agent

Audit every line, explain every finding

The audit agent runs a 25–40 turn adaptive-thinking loop against a tool layer that exposes rate cards, invoices, accessorials, and service guarantees. Every overcharge or claim-eligible late delivery comes back with the agent's full reasoning, stored permanently in agent_run_log for review.

  • Per-finding tool-call trail
  • Append-only audit log (Postgres trigger-enforced)
  • Human confirm-before-claim
Operator assistant

Ask, propose, confirm

An embedded chat assistant with read access to your shipments, invoices, rates, and findings — and write access guarded by confirm-before-execute workflows. Ask "tender these 40 LTL shipments at lowest cost" and the assistant proposes a plan, you review it, and the platform executes only on your confirm.

  • Permissioned tool catalog (MCP)
  • Per-tenant data scope, every call
  • Multi-step workflows with pause-for-confirmation
Anomaly detection

See what changed before you're asked

Continuous monitoring across spend, transit, accessorial mix, and on-time performance. When a carrier silently raises a surcharge or a lane's transit drifts, you see it on the dashboard before the next invoice arrives.

  • Per-carrier, per-lane, per-accessorial baselines
  • Threshold + statistical-drift triggers
  • SignalWatch alerts via Slack, email, webhook
Batch tendering decisions

Pick the right carrier for every shipment, every time

The tendering decision engine rates every shipment across every eligible mode and carrier, then applies a policy you control — cheapest, fastest, cheapest within transit, prefer mode, or prefer carrier. Constraints layer on top: lane blacklists, daily capacity caps, accessorial compatibility, service-level floors. Every decision writes an audit trail with both the engine's pick and any operator override — so when a finance review asks "why FedEx and not YRC", the answer is one query away.

Operator UI
Tender-board with proposal review + bulk dispatch
Agent surface
MCP tender_shipments tool with dry-run
Audit trail
Engine pick + override reason on every row
Architecture

How the AI is wired in

Three principles keep the agents useful and the operator in control.

1

Agents never touch the database directly

Every read and write goes through a Model Context Protocol (MCP) tool. The MCP server enforces tenant scope, validates inputs, and writes through the same service layer the operator UI uses. The model can't bypass row-level isolation or permissions — even if it wanted to.

2

Writes pause for confirmation

Multi-step workflows — batch tendering, claim filing, rate-card reparses — pause on a pause_for_confirmation() primitive. The agent renders a proposal; the operator confirms, edits, or cancels. Nothing irreversible happens autonomously.

3

Every step is auditable, forever

The agent_run_log table is append-only (Postgres trigger enforced). Every tool call, every reasoning step, every input and output is stored. When an audit finding is challenged six months later, you replay the exact chain that produced it.

Agent run detail showing turn-by-turn system, assistant, and tool result steps with token counts and timings

A real audit agent run — every turn labeled, every tool call captured, every token counted.

Models

Built on Claude

Toretto runs on Anthropic's Claude family — currently Claude Opus and Sonnet 4.x — chosen for tool-use reliability, document parsing strength, and adaptive thinking on long agentic loops. We don't lock you to a model: the agent layer abstracts the provider so we can move on price/performance shifts.

Prompt caching is enabled by default on every long-context tool invocation. Token spend per audit is tracked, attributed, and metered at the org level.

Your data

Stays yours

Your invoices, rate cards, shipment data, and findings are stored in your tenant — never used to train models. API calls to Claude run under Anthropic's no-training-on-API-data terms, and we never share your data across tenants.

On the audit side, all credentials (carrier portal logins, EDI VAN keys, telematics tokens) are stored with AES-256-GCM envelope encryption — never in plaintext, never exposed to the agent layer.

See an AI audit run on your own data

Bring a rate card and an invoice. We'll run the audit agent live in front of you — and you'll see every reasoning step, every tool call, and every finding the model surfaced.