One address, read by the people who build the thing: hello@atomelm.com.

What to put in the message

A question about a tiny model is only answerable with the constraints attached. The five lines that turn a vague enquiry into a measured answer:

The part. Exact MCU part number, or the two you are choosing between.
The budget. Flash and SRAM you can actually spare, in bytes, after the rest of the firmware.
The input. Sensor, sample rate, window length — or the text, and roughly how much of it.
The decision. What the device has to output, and how often.
The constraint that hurts. Battery life, certification, latency, unit cost, no-network. Usually only one of them is really binding.

With those five, the answer is a number and a yes or no. Without them it is an opinion.

Things worth asking

Whether a model of the size you need fits the part you have already chosen — see the measured RAM table first, it may answer it.
Whether the TensorFlow Lite for Microcontrollers comparison holds for your task. It is one dataset; running it on yours is a real question and a short piece of work.
Whether we can run the EDGE benchmark harness against your data. It is not in the public kit, and the answer is usually yes.
Whether you benchmarked Atome LM and it lost. Send it. It goes on the press page next to the results that flatter us.

Silicon bring-up, secure boot, per-platform hardening, regulated-device pilots and domain fine-tuning are described on the services page.