One path in
Every record — live feed or historical backfill — travels the same five stages. No side doors, no manual patches, no exceptions for urgent data.
Project brief · v2
Start over, properly. One pipeline, one model, both provable.
The legacy platform proved the idea and exposed the fault line: the data pipeline and the prediction model grew together until neither could be trusted or reproduced. ATORA keeps the understanding and the brand, and rebuilds those two systems from zero with contracts between every stage.
Rebuild principles
Contracts, not clean-up
Each stage publishes what it guarantees. When the guarantee breaks, the run fails loudly instead of quietly repairing the row and moving on.
Models read snapshots
The prediction layer never touches live tables. Every prediction cites the dated snapshot and feature version it was built from, so it can be reproduced exactly.
Scored against reality
Accuracy sits next to calibration error and sample size. A model stays in draft until it beats the deliberately simple baseline on a held-out window.
The pipeline, end to end
- 01
Ingest
- 02
Normalise
- 03
Validate
- 04
Feature build
- 05
Publish
Stages are separate, ordered and inspectable. A run that violates its stage contract stops there — it never reaches the model layer.
What carries over from the legacy platform
- Brand and design language
- Kept — cream, navy, brown, gold.
- Domain understanding
- Kept — what we learned about the problem is the real asset.
- Legacy data pipeline
- Rebuilt from zero.
- Legacy prediction model
- Rebuilt from zero, starting at a baseline.
- Overlapping feature modules
- Not carried over.