Fleet exception co-pilot
Edge-deployed assistant for dispatchers. Tool-calling with strict policy checks against live telemetry; written handover to an internal team.
- Venture-backed mobility
- Agents
- 6 weeks
- Tool calling
- Policy
- Edge deploy
- Runbook
Best fit when.
- 01Operators triage exceptions across multiple screens and want a single composable view, not a new system.
- 02Every action must already exist as a primitive in your stack — the agent composes, it does not introduce write paths.
- 03Telemetry can be stale and the surface must degrade visibly rather than guess.
What was happening.
Dispatchers triaged exceptions across a moving fleet — late arrivals, route deviations, vehicle faults. The previous tooling required them to pivot between four screens to assemble a single decision. The team wanted a co-pilot that could compose the picture and propose actions, while keeping the dispatcher firmly in the loop.
What we were holding to.
- Network was unreliable in the field; the surface had to degrade gracefully when telemetry was stale.
- Every action proposal had to map to an existing dispatch primitive — no new capabilities, just better composition.
- The dispatcher was always the deciding party; the co-pilot proposed, never executed silently.
How we built it.
Tool calls bounded to existing primitives
The agent surface called into the same dispatch APIs the human team already used. No new write paths were introduced; the model could only propose actions the existing system could already perform.
Telemetry-aware policy checks
Before any proposal surfaced to the dispatcher, a policy layer checked the live telemetry context — vehicle status, route window, regulatory cap. Stale telemetry collapsed the proposal to a 'gather more data' state instead of a confident recommendation.
Written handover with a cross-trained owner
The engagement closed with a written runbook and one cross-trained engineering owner inside the client team. We stayed embedded for one iteration cycle after handover, then left.
The shape sits exactly where we draw the line in our writing on when not to ship an agent — the model picks the route, the existing dispatch system carries the load, and no new write paths were introduced.
What we left with the client.
- Edge-deployed assistant integrated with the existing dispatch stack.
- Policy layer with telemetry-context checks and graceful-degradation modes.
- Runbook covering fallback behaviour, model swap procedure, and incident triage.
- Cross-trained engineering owner inside dispatch operations.