The control plane for models you own.
Philotic Desktop runs the Forge workflow. It keeps the record of how every model was taught, tested and released, and refuses the steps that would make that record untrue. The platform is our product. What it produces from your data is yours.
Philotic Desktop with three illustrative workspaces and sample data; none is a customer.
The Forge loop
Every pass makes your layer better.
- PackStart from a kind of work, not a blank page.A pack brings the label questions, training recipes and evaluation templates for a kind of work, such as inspection or video scoring. It carries no data, and applying one never overwrites what a project already has.
- ReviewPeople teach the model by answering.A model proposes an answer with its evidence; a person confirms, corrects, rejects or says what would settle it. When a newer model changes its proposal, the earlier decision is marked stale and comes back.
- DatasetAnswers become versioned data.Reviewed items publish as dataset versions with their lineage. The held-out split is locked by a stable hash, so an evaluation example never drifts into training.
- RunA run pins exactly what it learned from.A run pins its recipe, foundation, dataset version, exact training rows and evaluation set. A run whose training rows overlap its evaluation set is refused before it starts.
- EvaluateEvery result belongs to one version.A result is bound to one model version and one version of the evaluation set, with intervals, slices and item-by-item comparison. Change the evaluation set and earlier results stop counting.
- GatePromotion needs evidence and a reason.A candidate is promoted only with a fresh result on the current evaluation set, every required bar met, and a written note. It warns when any metric or slice is worse than the live version.
- ReleaseAn append-only history, with rollback.Promoting retires the live version and rollback restores the previous one, each recorded with its time and note. Every version records the run, recipe, data, foundation and license behind it.
- FlagProduction mistakes become the next lesson.Anyone can flag a production output, which goes back to review as a proposal. Once enough corrections collect since the last release, the project proposes a retrain.
From the last step the loop starts again: flagged outputs go back to review, become labels, and teach the next version.
A labeling team, not just a labeling tool.
Review is where your experts teach the model. Behind it, Philotic’s own team labels to one standard per domain, sends the hard calls to domain experts, and freezes the test sets every version is judged on.
Built for evidence, not for demos.
One record per project
Versioned label questions
Experiments on whole variants
Agents that propose, people who accept
Where the platform is going.
These are on our roadmap and not in Desktop yet. We list them so you know what we are building toward, and what is not there today.
Planned
Training dispatched from the record
Launch a run on your own cloud account or servers, or on ours, from the same recipe record, with progress reported back. Today training runs on GPUs outside the app and its result is recorded against the run.
Planned
Roles and two-person sign-off
Sign-in and roles, so labeling, review, promotion and rollback can be assigned to different people, and a release can require a second approver.
Planned
Deploy from the same screen
Push a promoted version into your environment from Desktop. Today promotion is the record of what is live.
Planned
The owner’s record, enforced
The Ledger shows who owns each model, the license of every component and where each runs, with a readiness checklist. Filling it in from every project and making the release gate enforce it are next.
Planned
Traces and cost
Run timelines, production traces, and cost by call, run and project.
Retraining is the work we keep doing.
The pipeline stays with Philotic, the way a software vendor keeps its product. You get the models it produces from your data, a license to run them where you choose for the term, and the loop that keeps them improving: new labels, new evaluations, and new versions through your gate.
Build the model your work needs.
Tell us what your people inspect, grade or decide every day. We’ll test candidate foundations on your task and show you the results before anything goes live.