🚜 Solutions

Heavy Equipment Intelligence

An in-cab assistant that runs on the compute OEMs already ship: spoken control of work limits, machine settings, and diagnostics, plus a page-cited operator's manual — with every physical change confirmed before it takes effect.

16
In-cab machine functions
100%
Physical changes confirmation-gated
0
Cloud calls required at runtime

Live demo

Drive the cockpit yourself

A simulated excavator with live machine state, the full confirmation gate, page-cited manual answers, and a spoken voice mode. Every tool call is shown as it executes.

1.2BFunction-calling model
350MManual retrieval model
15MOn-device voice

Deployment envelope

In-cab compute
  • Confirmation-gated physical changes, enforced at runtime
  • Page-cited manual answers, zero generation
  • Runs offline on in-cab compute

Operators lose seconds to menu trees all shift long, and new operators wait on supervisors for answers the manual already has. A compact function-calling model routes operator intent to machine functions against live machine state; work-limit and control changes are confirmation-gated by a runtime safety contract, the same way every time. Manual questions route to a retrieval model that serves vetted answers with page citations. Small enough for existing in-cab modules — no new hardware, no connectivity requirement, no per-query cloud cost.

3 specialist models

How It Works

Each model solves one hard problem,deployed where it matters

01

Hands on the Joysticks, Not in the Menus

Work area limits, machine settings, climate, lights, wipers, radio — every adjustment today costs a trip through nested screens. A fine-tuned function-calling LFM turns one spoken sentence into the right tool call against live machine state, including parallel commands and mid-request corrections. Numbers survive exactly as spoken: forty millimeters is 40, not 4.

02

A Safety Contract the Model Cannot Talk Its Way Around

Anything that changes how the machine moves — work limits, boom and swing speed, control patterns, work tools — requires the operator's confirmation before it executes. The gate is enforced at runtime from transcript evidence, not by model politeness: an unconfirmed gated call is converted into the confirmation question it should have asked. Declines never execute, stale confirmations expire, and excluded actions (starting the engine, moving the machine) are refused outright.

03

The Operator's Manual, Answered — Never Paraphrased

Maintenance intervals, fluid specs, fault-code meanings: a retrieval model serves vetted manual answers with page citations. Questions the manual doesn't cover get an honest 'not covered' rather than a guess — the assistant routes to the manual, it never writes it. New operators learn where the truth lives on page one.

04

Better AI on the Hardware You Already Ship

The whole stack is sized for existing in-cab compute: capability added by software, deployable across a global install base including sites with no coverage. Speech and machine data stay in the cab. A cloud assistant remains an option for the heavy tail — hybrid by choice, not by necessity — and the same footprint supports net-new features: in-vehicle analytics and edge agents with model, harness, and memory on device.

Try each model

All Demos

Ready to deploy in your environment?

Sub-50ms intelligencefor every critical decision.