Metis
The estimate comes out as a 90% interval, with coverage measured on the data.
Features
- A 90% confidence interval whose coverage was verified on data
- The most similar past projects, each with its weight on the estimate
- An indicative delivery plan derived from the interval
- A validation backtest against the client's own closed projects
- Runs locally, with no API keys and no hosted models
A software effort estimation engine. From a requirement and a handful of structural inputs it returns a median estimate, a calibrated 90% confidence interval and the closest projects in the history it was trained on.
multi-role internal platform, 14 screens, ERP integration
- the single number everyone asks for
- on its own it does not hold up against the data
- 90% interval, calibrated in log space
- coverage is measured on the data, not claimed
Metis estimates the hours a software project will take. The output is not a single number but an interval whose real coverage is known, because it was measured on data. Validation was pre-registered: the protocol was written and frozen before the experiments, across nine public datasets, with temporal splits and three baselines, including the human expert estimate wherever the dataset records one. Every outcome was published. On point accuracy the result is negative: on cross-organisation data the central value does not reach the threshold that was set. The confidence interval, on the other hand, held up. The product is built on that finding. The [p5, p95] band is calibrated in log space, so it stays balanced between small and large projects and its lower bound is always positive. The coverage the API reports is the coverage measured when the active model was fitted. Before anyone adopts it, a pilot runs a backtest against the client's own closed projects and compares it with the estimates their team produced at the time. The verdict is deterministic and has three outcomes: positive, negative, or inconclusive for lack of data.