Human Review & Continuous Learning
The model proposes, a person decides. Trusted moderators check every detection, fix the boxes it got wrong, and reject photos that should not be used. Their corrections are what the model learns from next.
A model that trains on its own guesses only gets more confidently wrong.Basis keeps a human between the model and its own training data, and lets the model learn most from exactly the cases it got wrong.
Step by step.
The AI proposes boxes
Each new report appears in a moderation queue with the model's boxes drawn over the photo.
A moderator reviews it
Approve the boxes as they are, drag and resize them, change the class, draw a missing one, or reject the photo entirely.
Reviewed photos are collected
Approved and corrected photos become training examples. A photo confirmed to contain nothing is kept as a useful negative example.
The model is retrained in batches
Once enough reviewed photos have built up, a fine-tuning run starts on cloud GPUs. The new model only replaces the live one if it scores at least as well on photos it has not trained on.
The details.
A purpose-built reviewer
A web reviewer with drag, resize, redraw and class-change tools. It works from any browser.
Several trusted moderators
Access is granted per person and every decision records who made it.
Nothing goes live blindly
New model versions are scored against held-out photos before they replace the old one, and every run is logged.
Wrong boxes become the lesson
When a moderator corrects a box, the correction is the training signal.
- Review
- Approve, correct or reject
- Trigger
- A batch of reviewed photos (about 1,000)
- Safeguard
- New model must match or beat the old one
- Audit trail
- Every run and decision is recorded
Help review
Moderator access is granted by invitation.