Use cases
The same engine, contract and scoring in every field. Each example links to a real test on public data.
Life science and lab data
Labs, biotech, clinical research
Classify samples from protein panels, sequences or clinical measurements, and cell images. Split by experiment, batch or patient, so the score reflects a new experiment, not a near-copy of the training data.
Containers teams build
- A protein-panel classifier for treatment groups
- Gene-boundary detection in DNA sequences
- Cell-type classification from microscope images (CNN on GPU workers)
- Risk flags from clinical measurements, with a written “not for diagnosis” limit in the contract
Tested on public data
- Mouse brain proteins: 35.7% right (+31% fewer mistakes than a simple guess)
- DNA splice junctions: 96.7% right (+93% fewer mistakes than a simple guess)
- Blood cell images: 81.2% right (+79% fewer mistakes than a simple guess)
- Heart disease (Cleveland): 82.0% right (+56% fewer mistakes than a simple guess)
Markets and commodities
Research desks, treasury, commodity buyers
Forecast forward returns with ranges (q10, q50, q90), tested walk-forward with a gap, and detect market regimes. Results come with the plain finding that short-term prices are close to unpredictable; the regimes are usually the more useful output.
Containers teams build
- 20-day forward return ranges for a commodity or index
- Calm / stressed / crisis regime labels as an input to other models
- Import public series from FRED by ID
Tested on public data
- Brent crude oil, 20 days ahead: 10 percentage points average miss (-1% smaller average miss than a simple guess)
- Natural gas (Henry Hub), 20 days ahead: 19 percentage points average miss (+1% smaller average miss than a simple guess)
Manufacturing and operations
Plants, fabs, logistics, product teams
Predict pass/fail and defects from sensor readings, forecast demand by the hour, and model how machines or customers move between states.
Containers teams build
- Yield and defect prediction from process sensors
- Hourly demand forecasting with a time split
- Next-state prediction for machine status or customer journeys (Markov chains)
- Customer or applicant risk scoring
Tested on public data
- Bike sharing demand: 43 rentals an hour average miss (+76% smaller average miss than a simple guess)
- Adult census income: F1 80 (+66% fewer mistakes than a simple guess)
- German credit risk: F1 68 (+45% fewer mistakes than a simple guess)
- California housing: $29,400 average miss (+67% smaller average miss than a simple guess)
Have a different question?
If it can be written as a contract (what goes in, what comes out, how it’s scored), it can be a container.
Request early access