Edge AI: the model runs here, not in a cloud
A camera that has to label a frame with no network. A robot that has to decide the next move when the radio is down. A battery-powered node that reads sensors and acts — sleep, wake, alarm, move — without sending the raw data anywhere. Olivabot B.V. (Amsterdam) builds that. Remote, autonomous, email first. Outline, then milestones you can check, pay, and continue — or stop. Hardest risks first. Email services@olivabot.com.
How we charge
No day rate, no hourly rate, no timesheet. Email, then a project outline, then we work in milestones — remotely and autonomously. We say up front what we think it will take, and where it might go wrong, and start with the hardest problems. Each step is quoted in writing. You check the work, pay, and decide whether to continue. A call if you ask.
What this actually is
- Image recognition on the device. A compact camera or an existing sensor. Labels, a confidence score, and an explicit “unclear” — not a round trip to a US API.
- Local reasoning for robots. The loop that must keep running when the link drops: what the sensors say, what that means, what to do next. Coordination with motion is firmware; this page is the on-board decision.
- Battery-powered AI on sensors. Classify soil, motion, light, gas, or whatever the probe already measures, then act: log a rare event, wake the radio, stay asleep. The model is part of the power budget, not an afterthought.
Tiny models on microcontrollers (Cortex-M, ESP32, nRF), slightly larger ones on a camera module or a small NPU next to the board. C, C++, and whatever runtime the chip already has. Quantised. Measured on the hardware you will ship, not on a laptop.
What it is not
It is not a ChatGPT wrapper and not a model that phones home for every frame. Photos that can live on a dedicated EU server are a different job — useful when you have a network and do not need an answer in milliseconds on a battery. We will say which of the two you actually have.
We do not sell cameras, NPUs, or robots. We deliver code and, if needed, CAD. You flash the board. Gibbon Bot is our own experimental prototype and does not work yet; it is not a product and not for sale yet.
Honest about the model
A model is only as good as the labelled data you have, and only as small as the chip and the battery allow. If it will not fit, or if there is not enough data to make something useful, we say that up front — not after you have paid for a training run. We ship confidence and “unclear” rather than inventing precision.
Where this has been proven
On-device movement classification on a finger-ring sensor for Madglove — battery, microcontroller, and a model small enough to sit on a finger. A feasibility study on on-board canopy vision without a cloud connection (the robot it was meant for does not climb yet). The same battery-and-chip discipline as the outdoor LoRaWAN probes for Mycofarming.
How to start
Email the device, the sensors or camera, and what it should recognise or do. Two or three sentences is enough — “ESP32-S3, a cheap camera, has to say fruit / leaf / unclear, no cloud, CR123A” is a perfect brief.
Reply within two working days. We work remote and autonomously, hardest risks first. A call or a video call if you ask. If it will not run on that battery, we will tell you that instead.
To services@olivabot.com. Under NDA if you prefer — just ask first.