Company

Exomantle

When software crashes, an engineer gets a stack trace, a stack of similar reports, and a link to the commit that caused it. When a warehouse robot stops in an aisle, someone walks over and looks at it. We are building the first thing for the second case.

Who is building this.

We came out of neural decoding research, where the problem every day was the same one: a raw signal stream too large to store and too noisy to read straight through. Robots have that problem with a different signal.

We built the on-robot agent before anything else, in Rust, and we ruled out two products early. We are not recording continuously into the cloud, because that is storage and other companies sell it. We are not writing our own file format, because we tried for a week and MCAP was better than what we had. What is left after those two decisions is Exomantle: the recording is chosen on the robot, captures from different machines with the same cause are grouped into one issue, and the replay opens in a browser.

We are two people, based in California.

Sandesh Bhandari

The agent and the data path

Before this, plant data systems at Norplex-Micarta.

Aamod Paudel

The cloud side

Before this, edge agent software at Cloudflare.

One of us comes from a factory, one from cloud infrastructure.

Where we are.

Playback ships today. The agent runs on ROS 2 Humble and Jazzy over Cyclone DDS and Fast DDS, on arm64 and x86_64 Linux. Re-running a single program against its recorded inputs, and coordinated capture across nearby robots, are both being built.

We are looking for fleets to run this on. If you have twenty or more robots in production and a failure you never fully explained, that is the conversation we want.

Talk to a founder

Hiring.

Two roles, our first engineering hires. Each one owns a system rather than a queue of tickets, and you will be on calls with the people running the fleets.

We are in California. Both roles are remote within US time zones.

Which topics carry a diagnosis and which are noise, what conditions should freeze a window, and how a 41 MB file gets off a robot on a warehouse link that drops. Native ROS 2, MCAP as the file format, and ros2_tracing for instrumentation light enough to leave on in production.

What we look for

ROS 2 experience on robots that ship to customers, rather than in simulation
Sensor data at volume, meaning video and point clouds rather than text logs
You have debugged a machine you could not walk over to

Both roles: $160k to $200k base, 0.5% to 1.0% equity on a four-year vest with a one-year cliff.

Apply: thirty minutes with a founder

The rolling 60 second buffer, the write path, and the C and Python wrappers around the SDK. It is Rust because a garbage collector pausing next to a control loop is not acceptable on a robot that moves near people, and because the memory bugs would be ours to find in production.

What we look for

Systems Rust rather than application Rust
You think in allocations and latency budgets
Embedded or realtime work helps

Both roles: $160k to $200k base, 0.5% to 1.0% equity on a four-year vest with a one-year cliff.

Apply: thirty minutes with a founder