Agentic Intelligence

The freight brain powering every decision.

Proprietary reasoning models trained on $2.3B of live freight operations — lane dynamics, carrier behavior, shipper patterns, and the messy edge cases general LLMs miss.

Why purpose-built matters

A general LLM can't tell a detention claim from a delivery confirmation.

Freight is a language — of lane codes, service levels, BOL layouts, EDI fragments, and dispatcher shorthand. Veltrum's models speak it natively.

Lane dynamics

Models the real-time price elasticity, capacity shifts, and seasonal patterns of 42K U.S. lanes and 120 cross-border corridors.

Carrier behavior

Predicts tender acceptance, on-time risk, and fallout probability per carrier per lane per day — not just as a static score.

Exception prediction

Flags shipments 4–6 hours before they break, with 87% precision. Time the agent has to act, not just alert.

Document understanding

Reads BOLs, rate cons, accessorial sheets, and POD scans — even handwritten driver notes — into structured facts.

Intent recognition

Understands dispatcher shorthand, broker slang, and shipper urgency cues — across email, chat, and voice.

Policy reasoning

Your margin floors, carrier blacklists, and SLA commitments become constraints the agent physically cannot violate.

Model stack

Three layers. One decision loop.

Veltrum Core

Freight-tuned foundation model. 34B parameters. Trained on operational logs, EDI corpuses, and dispatcher language.

Veltrum Predict

Specialized models for ETAs, carrier acceptance, exception risk, and lane price. Updated hourly from live telemetry.

Veltrum Policy

Your org's constraints and objectives compiled into a verifiable policy layer. The agent plans within it — always.

Transparency

Every decision, explained.

Autonomy without observability is unacceptable for enterprise ops. Veltrum ships reasoning traces by default.

TRACE

Readable reasoning

Every input, weight, policy, and alternative — rendered in plain language and structured data.

CONFIDENCE

Calibrated scores

Each decision is tagged with a confidence band. Low confidence auto-escalates; high confidence runs silent.

REVERSE

One-click rollback

Any autonomous action can be reversed. The trace is attached to the rollback for retraining.

Under the hood

Want the model card?

Our team will walk you through training data, evaluation benchmarks, and the exact guarantees behind every autonomous decision.