MimirMIMIRRequest a demo

03 — Status

Early, and saying so.

Most companies at this stage publish a roadmap and let you assume it's already built. This page is the opposite: what runs today, what got built and when, and a list of the things that genuinely aren't done.

  1. 01

    Collecting the field set

    We're out flying missions and recording real sensor footage — across terrain, altitude, weather and light. Public datasets don't look anything like the field. Ours does.

    In progress
  2. 02

    Labelling and training

    Every frame we capture gets annotated and folded back into the detection models. The set grows, the models sharpen, the loop repeats.

    In progress
  3. 03

    Onboard runtime

    Compressing perception down to what the operator already carries — real-time inference on the controller, with no link back to a server.

    Up next
  4. 04

    Public demo

    The first live look at Mimir resolving a real scene. We'll share it here, and send it to everyone on the list below.

    Soon

03.1

What works today

Everything below has run end to end on real hardware — not in a simulator, and not on one hand-picked video file.

Detector
An anchor-free centre-point detector, roughly 6.4 million parameters, output stride 4, single class: person. Trained from random initialisation on frames we captured and labelled ourselves, exported to ONNX and run locally.
Live feed
Video from a stock aircraft and controller, landing on a local media server and read straight into the workspace as a live tile. Detection and tracking treat it identically to a file — they cannot tell the difference.
Capture loop
Frame harvesting that runs whether or not a model exists, with a scene-change test in front of it so a static hover does not flood the disk. Every kept frame is written with its labels beside it.
Acceleration
GPU inference where the machine has one, automatic CPU fallback where it does not. No configuration, no separate build.
Licence posture
Nothing copyleft in the shipping binary. Training tooling never leaves the training machine; only the exported model file crosses into the product.

03.2

The build log

The dated version, for anyone who would rather see the shape of the work than a summary of it.

  1. June 2026

    First working prototype

    A desktop workspace that loads aerial imagery, runs local segmentation and detection, and puts results in a side panel. Enough to prove the interaction was right — and to show what was wrong underneath it.

  2. August 2026

    Data collection goes continuous

    Capture moved out of a manual step and into the app itself, so a session of use produces labelled frames instead of only output. The dataset stopped being something we made time for and became a by-product of the work.

  3. August 2026

    Clean-slate rebuild of the detection stack

    The whole detection layer was torn out and rebuilt so nothing in the shipping product carries a licence that reaches the weights. It cost weeks. It is also not something you can retrofit later, which is exactly why we did it early.

  4. August 2026

    First detector trained from scratch

    120 epochs from random initialisation on our own labelled frames, exported to ONNX with a documented decode. No pretrained checkpoint anywhere in its history.

  5. August 2026

    Live flight video, end to end

    Video off a stock controller and into the workspace as a live tile, verified on real hardware over a real link. Detection, tracking and capture all worked on it unchanged — pixels are pixels.

  6. September 2026

    Field data collection, ongoing

    Where we are now: flying, capturing, labelling, retraining. The bottleneck is honest coverage across terrain, altitude and light, and the only way through it is hours in the air.

03.3

What isn't done

The useful half of a status page. If any of this is a blocker for you, better that you know it now than after a call.

No public demo yet
It is the next thing we ship, and everyone on the early-access list gets it the day it exists.
No published accuracy
We have numbers. We are not publishing them, because the validation set is not yet broad enough across terrain and altitude for them to say anything about footage you would actually shoot. Quoting them now would be marketing, not measurement.
One class
People only. Vehicles and structures are next, and they wait on data rather than on architecture.
Not yet on the controller
Inference runs on a machine near the operator, not on the controller itself. The on-device build is written and waiting on hardware and toolchain — not on an unanswered design question.
No pricing
There is nothing to sell yet. When there is, it will say so here first.

Follow it from here.

One email when the demo lands. No newsletter, no drip sequence.

Get the demo→Why we build it this way→