01
Amazon SageMaker AI
Fincount AI trains its own counting, sizing and quality models on Amazon SageMaker AI, and tests each new version before it is released.
Technology
Built on AWS.
Fincount AI is trained and checked on AWS. A small version runs on the phone, so counting works with no signal. Videos are kept with a fingerprint, certificates are signed, and your data is stored in the AWS Africa (Cape Town) Region.
The six pieces
What each part does for you.
02
AI on the phone with ONNX Runtime
A small version of Fincount AI runs on the phone with ONNX Runtime. Counting works offline: the first count needs no signal and takes seconds.
03
The cloud check on AWS
Every basin you upload is recounted on AWS by a stronger version of Fincount AI. That recount is the final count. If it and the phone disagree, a person looks.
04
Amazon S3
Videos are kept in Amazon S3 with a fingerprint of each file, so a changed video shows. Only short lived links can read them.
05
AWS KMS
Every count certificate is signed with a key held in AWS KMS. The key never leaves it, and any change to the count breaks the signature, so a buyer can check it.
06
AWS Africa (Cape Town) Region
Your videos, counts and certificates are kept in the AWS Africa (Cape Town) Region.
How the pieces fit
From the basin to the buyer.
01
Phone
Records 10 seconds and gives the first count with ONNX Runtime.
02
Amazon S3
Keeps the video with a fingerprint. If the signal drops, the phone sends it again by itself.
03
Cloud check
Fincount AI recounts and measures every uploaded basin on AWS.
04
AWS KMS
Signs the sale so any change shows.
05
Buyer
Scans the QR code and checks the count as issued.
- 01
Phone
Records 10 seconds and gives the first count with ONNX Runtime. - 02
Amazon S3
Keeps the video with a fingerprint. If the signal drops, the phone sends it again by itself. - 03
Cloud check
Fincount AI recounts and measures every uploaded basin on AWS. - 04
AWS KMS
Signs the sale so any change shows. - 05
Buyer
Scans the QR code and checks the count as issued.
Open source parts
What Fincount is made of.
Every part has a licence that allows commercial use. We do not use code under licences that would force us to publish Fincount's own code.
| Part | What we use | Licence |
|---|---|---|
| 01Counting models | Our own models, trained by us from our own data | PyTorch, BSD style licence |
| 02Image tools | OpenCV, NumPy, SciPy, scikit-image | Apache 2.0, BSD 3 clause |
| 03Phone | Kotlin, CameraX, WorkManager, Room, Media3, ONNX Runtime | Apache 2.0, MIT |
| 04Server and web | FastAPI, SQLAlchemy, Alembic, PostgreSQL, Next.js | MIT, PostgreSQL licence |
| 05Certificate PDF and QR code | ReportLab and segno | BSD |
Count your next basin with Fincount.
Film a basin, get the count, and send your buyer a count they can check. The first hatcheries get 30 days free.