Telemetry & Ops
Health, latency, intent logs, alerts, fleet view, training feedback loop. The dashboard a real on-call engineer watches at 2am when something is off.
- Prometheus
- Grafana
- Sentry
- Custom audit log
- PagerDuty / Opsgenie hooks
Five layers under every robot we ship.
We document the architecture because the buyers who matter ask. If your CTO wants a deeper read or a custom variant, write us. We will send the whitepaper and a sample repo.
Agent Layer
Task graph, brand voice, persona, safety envelope, escalation paths, handoff to humans. This is the layer that makes a robot feel like a coworker instead of a parrot.
- Custom Python orchestrator
- LangGraph patterns where they fit
- Pydantic schemas
- YAML personas
- Eval harness with graded test set
Models
Whatever model wins for the task at hand. Long-context reasoning here, fast tool-calling there, on-device speech for latency, on-device LLM when the network is unreliable.
- Claude 4 family for reasoning
- GPT-4o for tool calling
- Llama 3 8B Q4 on Orin (fallback)
- Whisper-small.en (ASR)
- Piper / ElevenLabs (TTS)
- Moondream 2 (VLM)
Orchestration
The wiring between the SDK and the agent. Message bus, task scheduler, sensor fusion, the bits that ROS 2 does well and the bits we replace because ROS 2 does not.
- ROS 2 Humble
- micro-ROS bridge
- DDS
- Custom Python services
- Redis for ephemeral state
- PostgreSQL for episodic memory
Vendor SDK
The hardware floor. Manufacturer SDK, low-level motor control, sensor I/O, on-robot networking, kill switch, charging dock protocol.
- Unitree SDK 2 (low + high level)
- Boston Dynamics SDK
- Figure / 1X partner APIs
- CAN bus + GPIO for retrofits
Five rules we will not break.
- Models change. Interfaces should not. The agent layer talks to a model interface, not a vendor SDK. Swapping Claude for a local Llama is a config change, not a refactor.
- Every agent has an escalation path. If the agent cannot answer with a confidence threshold, it hands off to a human with structured context. No silent failures.
- Evals are not optional. Every deployment ships with a graded eval set. Regressions break the build, not the customer experience.
- Telemetry is a first-class citizen. If we cannot see what happened in production, we did not deploy a system, we deployed a prayer.
- The kill switch is hardware, not software. Anything that moves around humans gets a physical, visible, single-action stop. No exceptions.
Want the deeper read?