Comprehensible

When AI operates in the real world, the stakes are high

Reliability for
Physical AI

How soon will you know
when your robot fails?
How fast can you recover?

We built reliability solutions at world-leading companies.
Now we’re bringing the same principles and technology to physical AI —
adapted to the scale and demands of the real world.

Explore the product

Move fast. Don’t learn
the hard way.

Catch rare failures early — because one serious incident is one too many.

The Product

Finding issues starts with defining nominal behavior

Real-world complexity makes this hard. Even something
as intuitive as “a left turn” requires defining how
position, orientation, velocity and acceleration relate
over time — including where the turn starts and ends. We built a solution for this.

Specification

Reliability needs answers you can reproduce, explain and verify

Engineers describe expected behavior in plain language. An AI assistant helps translate it into a specification: a precise, reviewable definition that can be applied to recorded and live data.

A specification can capture patterns across time and space, statistical methods, anomaly detection and learned dynamics.

The system evaluates recorded or live data against the specification without relying on AI, so the same data always produces the same result. Every finding links back to the relevant parts of the specification and source data, showing engineers what was detected and why.

  • Compose simple definitions into complex behaviors
  • Inspect every intermediate definition and relationship
  • Validate specifications before execution
  • Review changes as a normal code diff

From simple to complex

  1. 01Sensor readings
  2. 02Relationships
  3. 03Behaviors
  4. 04Scenarios
Example Prompt

Define nominal flight-control behavior: measured roll, pitch and yaw should follow their commanded values over time.

Reviewable specification 
[metric.causality:control]
roll_response = roll.commanded -> roll.measured
pitch_response = pitch.commanded -> pitch.measured
yaw_response = yaw.commanded -> yaw.measured

[spec:control_nominal]
control.roll_response
control.pitch_response
control.yaw_response

Real-world UAV monitoring · Discovering causal relationships and combining them into nominal control behavior

Example Prompt

Flag an interrupted motion when an unusual change in motor load is followed by the robot stopping before it reaches its target.

Reviewable specification 
[event.anomaly:motor_load]
robot.motor.load

[event:motion]
started = robot.motion.active
incomplete = robot.motion.stopped && !robot.target.reached

[event.sequence.30s:interrupted_motion]
motion.started -> motor_load -> motion.incomplete

Robot motion monitoring · Combining anomaly detection, state and sequence into one explainable failure condition

Testing, verification and monitoring

From regression testing to real-time monitoring

Use specifications to test recorded runs, catch regressions after every commit, verify edge and tail cases, and continuously monitor live systems. Developers can move quickly without changing what “correct” means between development and operation.

Verification
Checks the system against its specification — not just familiar test scenarios — including edge conditions and low-probability tail cases.
Anomaly detection
Rare failures provide too few examples to learn from. Anomaly detection finds behavior that differs significantly from normal without requiring every failure mode to be known in advance.

Time, spatial, interval and event indexes narrow each search to the relevant data, keeping it fast as datasets grow. Distribution-shift detection identifies when the statistical behavior of data changes over time. Our optimized Kolmogorov–Smirnov test provides a transparent, low-latency alternative to model-heavy approaches.

01 Testing Check recorded runs
02 Regression Recheck every commit
03 Verification Probe edge and tail cases
04 Monitoring Evaluate live systems continuously

Measured CPU distribution-shift detection

64.2B events per second Memory · uncompressed stream · CPU-bound
5.33B events per second SSD · Iceberg · I/O-bound
533.3M events per second AWS S3 · Iceberg · network-bound

12-core Intel i7

4 μs Our implementation
vs.
2,000 μs scikit-learn

Kolmogorov–Smirnov test · 10K samples · identical hardware · 500× measured speedup

GPU-accelerated indexes · Time · Space · Intervals · Events

Investigation and response

See what happened. Understand why. Recover safely.

Comprehensible shows what happened, why it was flagged and which expectation it violated. It generates an incident-specific view from the query and specification, then helps teams trigger an automatic or human-approved response and confirm recovery.

  • See what the machine sensed and did before, during and after an incident
  • Compare what happened with expected behavior and similar past runs
  • Turn what the team learned into regression tests and real-time monitoring
  • Use existing recovery systems to respond automatically or with human approval
Time-series diagnostics for robot joint position, velocity and torque, followed by a friction profile
Time-series diagnostics · Position, velocity and torque
Feature-extracted friction model showing raw telemetry, binned medians, a sliding friction fit and stiction peaks
Evidence to model · Extracted friction behavior
Robot telemetry being investigated in Foxglove on a workstation
Field investigation · Foxglove

Team interoperability

One source of truth. Familiar tools for every team.

Reliability crosses development, field operations, support and leadership. Comprehensible gives each team a familiar interface while keeping everyone aligned on what should happen, what actually happened, what was decided and whether the system recovered.

  • Write specifications and investigate in the Comprehensible VS Code extension, with MCP-assisted authoring
  • Review sensor data and robot state on the same timeline in Foxglove or Rerun
  • Share incident status, evidence and actions in a team view

Leadership

Built by people who have led
reliability at global scale.

Bob Nugman

Co-founder & CEO

Bob Nugman

Built reliability systems and validation programs for Uber’s core trip flow, DoorDash’s production and ML systems, and Shield AI’s autonomy stack.

Alex Kabirov

Co-founder & CTO

Alex Kabirov

Led large-scale reliability projects for Fortune 500 companies and other leading global enterprises.

Build with us

Physical AI reliability. Made comprehensible.

Bring us the behavior your machines must get right. We’ll help you define it, test and verify it, monitor it in real time, and recover safely when behavior changes unexpectedly.

hello@comprehensible.ai