Safety-first AI for autonomous drones

Autonomy has layers.

Guru is the AI software layer for autonomous drones — a multimodal brain that understands your mission, wrapped around a deterministic safety kernel that never guesses. The model advises. It never touches the motors.

Request a demo Explore the layers

independent layers, from senses to safety kernel: 3; independent layers, from senses to safety kernel
chain of command from mission to motors: 1; chain of command from mission to motors
of decisions recorded and replayable: 100%; of decisions recorded and replayable

The layers

Three layers. One chain of command.

Most autonomy stacks let a neural network fly the aircraft and hope for the best. Guru separates intelligence from authority: the AI understands and proposes, and a deterministic kernel decides what actually happens.

Layer 01

The Senses

Voice, vision, and live telemetry fused into one picture of the mission — so the aircraft understands what it hears and sees, not just where it is.

Layer 02

The Brain

A multimodal AI that reasons about the mission and proposes the next action as a strict, versioned contract. Advisory by design — it can suggest, never seize.

Layer 03

The Kernel

A deterministic supervisor, written in Rust, with sole authority to act. Every proposal is checked against geofence, battery, and vehicle state — and the emergency stop latches until a human clears it.

Fails safe

Lose the model, lose the link — the kernel notices on its own and puts the aircraft in a safe state.

Black-box evidence

Every request, decision, and action is recorded and correlated, so any flight can be replayed and explained.

Any airframe

Replaceable vehicle adapters keep the intelligence portable across drone bodies and flight controllers.

Human override

Arming and takeoff always require explicit confirmation, and a hard stop outranks everything.

How it works

The model proposes.
The kernel decides.

Every action the AI wants to take travels through one narrow, auditable path — and the deterministic kernel is the only thing that can touch the vehicle.

1
Understand
Spoken instructions, imagery, and telemetry become structured mission intent — not free-form text steering an aircraft.
2
Propose
The brain answers with a strict, versioned action proposal. Anything malformed or unexpected is rejected outright.
3
Decide & fly
The kernel validates against geofence, battery, and vehicle state, then executes — or refuses. Critical actions wait for a human yes.

brain → kernel · action proposal

{
  "schema_version": "v1",
  "request_id": "9f2c41d8",
  "action": "take_off",
  "params": { "altitude_m": 30 },
  "rationale": "Begin tower sweep at inspection altitude.",
  "confidence": 0.92
}

✓geofence— inside mission boundary

✓battery— 87% — mission margin ok

✓vehicle state— armed and ready

⏸ holding for operator confirmation — takeoff never runs on model say-so

Use cases

Inspection, first.

Guru is built for work where a missed defect costs more than the flight — and a bad decision costs more than the drone.

Energy

Wind turbines and solar farms swept on a schedule, with every observation tied to evidence you can replay.

Towers & telecom

Close-in structural sweeps at inspection altitude — without putting a climber on the steel.

Pipelines & utilities

Long linear corridors patrolled autonomously, anomalies flagged with location, imagery, and severity.

Bridges & infrastructure

Repeatable flights of the same structure over months, built for comparison — what changed, and how fast.

Industrial facilities

Flare stacks, roofs, and tank farms inspected without scaffolding, shutdowns, or guesswork.

Environmental monitoring

Wildfire watch, coastal erosion, and habitat surveys flown by aircraft that explain what they saw.

Contact

Put a brain on board.

Tell us about your aircraft and your mission — we'll show you what Guru can do with them.

Request a demo hello@onion.guru