Before Exomantle existed, we worked on brain implants. The problem was reading intent from brain signals. A brain implant array has 10,000 channels. Each one samples 30,000 times per second. That is about a gigabyte of raw voltage data per minute, and most of it is noise.
The useful signal is weaker than the thermal noise, muscle movement, and electrical interference. An implant cannot run a model on itself. It does not have enough power, and it would get too hot.
So the processing moves off the device. You filter the signal, keep the spikes that carry information, match them to learned patterns, and return an intent. It returns a character, a word, or a cursor movement.
The whole loop has to finish in under 100 milliseconds or the interaction breaks.
10,000
channels on a dense intracortical array
30 kHz
sampling rate per channel
GB / min
of raw voltage, mostly noise
This was early research. We are not a medical device company and we do not sell neural decoders.
Then we looked at robots.
A robot's message stream produces 100 to 300 MB per minute. When something fails, the explanation is hidden in that stream. The robot cannot send everything to the cloud, and it cannot process it locally either.
We had already built the pipeline for this. You filter out environmental noise, keep only what carries useful information, match it to patterns from past runs, and return something a human can act on. The signal was different, but the question was the same.
Neither an implant nor a robot can send everything. Deciding what matters has to happen where the signal is.
Gigabytes reduce to a signature you can group and compare.
The limit is under 100 milliseconds for a conversation or a replay. After that, people stop opening it.
We do not sell a decoder. Exomantle is not a medical device company and has no clinical program. The neural work is where the thinking started.
We ship an agent that keeps the last 60 seconds of a robot's sensor data, and a replay of the minute before it broke. That is on the product page.
One idea connects the two. A decoder improves as it trains on more brain data. We think a failure library improves as it collects more data from more fleets, and that the patterns transfer across robot types. Nobody has measured that. We will run the experiment and publish what we find.