labs@atheron:~/work/aton-ai$ cat aton-ai.md
ATON AI
Trained on cards we own.
No cloud GPUs and no hosted API for the parts that matter. We train on our own cards, hold the latest months back, log every read of them, and keep a model only when it wins on data it has never seen. We publish no figure for how any of them performs.
// what it does
- Trains gradient-boosted trees on the GPU, with a neural network used only if it beats them on the same held-out data.
- Trains continually: one venue at a time, every interval from one minute to one week.
- Replays rules over history on the GPU, queued from the team portal and written back through a relay.
- Labels news with an open-weight model running locally: event, direction, urgency and topic, with no hosted API.
- Reviews our own code with automated checks that re-run the tests, part of the mutation set and secret scans every few hours.
// how we built it
- The latest six months are held out, and every read of them is logged, so a result cannot quietly be fitted to them.
- A governor holds the main card's load down while someone is working at the desk, and never leaves a job suspended.
- Training data is mirrored to a local drive and copied to fast storage for a run, so a run never waits on a network.
- Heavy work goes through one machine-wide lock, so two jobs never fight for the same card or the same disk.
- The team portal's own assistant runs on a frontier model and has no tools by design: it reads, drafts and summarises.
// the size of it
- hardware
- Our own cards: one for the models, one for the language model
- held out
- The latest six months, with every read logged
- intervals
- One minute to one week, venue by venue
- cloud GPUs
- None
// stack
- Python
- XGBoost
- PyTorch
- CUDA
- Hugging Face transformers
- TypeScript
// where it stands
Training is paused while the current build is finished. No model is released, no model trades real money, and we publish no figure for how any of them performs. The strategies and weights are private and stay private.
// a comparable build
- Build
- ≈ US$85,500 to US$131,000, within 20 weeksCAD 121,700 to 186,100
- Hosting
- ≈ US$520 a monthCAD 740 a month
- Running the AI
- ≈ US$176 a monthCAD 250 a month
- Support
- ≈ US$2,500 a monthCAD 3,565 a month
- Hardware
- ≈ US$47,800 to US$197,000CAD 68,000 to 280,800
Prices in your currency are estimates from today's Bank of Canada rate. All invoicing is in CAD or USD.
the effort in it
- Development
- 405 to 620 h
- Machine learning engineer
- 171 to 262 h
- DevOps engineer
- 86 to 132 h
- Senior engineer
- 79 to 121 h
- QA engineer
- 58 to 88 h
- In all
- 979 to 1501 h
what this price covers
A training and evaluation platform as one build: the data pipeline, the GPU training and backtesting, the evaluation harness, and the local model serving. The GPU hardware is priced with it and shown on its own line.
An estimate for a build like this one, worked out from our rate card as you read it. It is not a quote. Yours changes with your scope, and we write a real one after a call.
Open this in the estimator →// the rest
The other projects.
// start
Want one built like this?
The same team, the same tests, the same care. Tell us what you need, or price it yourself first.