Vane

A plan built from single-point estimates is wrong in a predictable direction. Tasks overrun more often, and by more, than they finish early. When several tasks queue up for the same person, those overruns don't average out. They stack. A schedule where every task is "probably on time" can be very unlikely to finish on time overall, and a spreadsheet has no way to show it.

The other blind spot is people. A spreadsheet shows that the team has 400 hours free next month. It doesn't show that 300 of the hours needed can only come from the two people who know the structural model.

Vane imports timesheets, task lists and contract milestones from the tools a firm already uses. From that history it learns how estimates have actually turned out, by type of work and by person. It fits a duration distribution to each task instead of a single figure, and records which skills each task needs and who can supply them.

Each simulation run then plays the schedule forward: it samples a duration for every task, respects dependencies and each person's real availability, and records when every milestone lands. Ten thousand runs give a distribution for every deliverable. Sensitivity analysis over the same runs ranks which tasks and people, if they moved, would shift the date the most.

Monte Carlo is random by design, and a forecast that changes each time someone presses refresh is useless in a meeting. Every Vane run is seeded and versioned against a snapshot of the input data, so the same inputs always give exactly the same forecast. When a date moves, the change can be traced to the timesheet entries or plan edits that moved it, not to noise.

The engine is written in Rust and compiled to WebAssembly, so a full forecast runs on the user's own machine in the browser, fast enough to rerun live while someone drags a task or reassigns a person. Schedule and staffing data never has to leave the building to be modelled.

The question Vane is built to answer is "can we take this job?" Before a bid goes out, the new project can be dropped into the live forecast to show what it does to every existing commitment, and who would have to be hired, moved or freed up to make it fit. That turns a gut call by partners into a conversation about specific dates and specific people.

The simulation core and the reproducible forecasting pipeline are complete. The interface is in beta, with a forecast view, a bottleneck ranking and what-if editing. Current work is on importers for common timesheet and project tools, so a firm can get its first forecast in an afternoon rather than after a data migration.

Next is calibration reporting: going back over past forecasts and showing how often the real date landed inside the predicted range. That's the number that should decide whether anyone trusts the next one.