Skip to content

Your first AI pipeline run

At a glance

Go from an AI provider that's already synced to your first automatically-classified resource: wire it onto a project, upload data, watch the pipeline run, and approve the result.

  • Time: ~15 minutes once a provider exists; AI run time depends on collection size and hardware
  • Who: anyone who's set up their first research project and wants to see AI classification end-to-end
  • Prerequisites: a Classification Project with Resources to classify; an AI Provider already registered

Before you begin

This tutorial doesn't explain how providers get registered — it assumes that part is done and walks through what happens once it is. For the configuration details behind each step, follow the links.

Steps

1. Wire the provider onto your Classification Project

Open your Classification Project (Media classification → Classification projects) and set:

  • Object detection AI model — a detection-type provider, e.g. MegaDetector v6
  • Species AI model (optional) — a classification-type provider, e.g. DeepFaune v1.3, runs only on objects the detector marked as animal

Save. From this point on, every newly uploaded Resource in the project's Collections is submitted to the AI Manager automatically — no manual trigger needed.

Full field reference: Create a classification project.

2. Upload (or re-use) some resources

If you already uploaded resources while following Your first research project, skip ahead — they'll start processing as soon as step 1 is saved, picked up by the next Celery beat tick. Otherwise upload a small batch now via the web UI or trapper-tools.

3. Watch the job run

Each submission becomes an AI Classification Job (Media classification → AI classification jobs). Open the Celery Flower dashboard to watch the underlying task move through the queue, or just refresh the job admin page — the Stage column updates as it progresses through the AI Manager → AI Worker → back to Expert round trip.

AI Classification Job admin list showing job stage and provider

A job that never leaves an early stage, or fails, is covered in Run & re-run the AI pipeline.

4. Review the AI Classification

Once the job finishes, open one of the processed resources in the classify view. You'll see an AI Classification already populated: bounding boxes, species (if a species model was configured), and confidence scores — all read-only at this point, this is the AI's untouched output.

Classify view showing an AI-generated bounding box around a detected animal, with confidence score and an Approve AI button

If nothing shows up here, check three things before assuming it's broken: the job's Stage in the admin (it may simply not be finished yet); whether the AI genuinely found nothing on this particular resource (a real, valid "blank" result, not an error); and — if you can't classify anything yet — whether the project's Required AI classification setting is on, which blocks manual classification until the AI Classification Job for that resource has completed.

5. Edit (fork) or approve directly

Two paths from here, both valid:

If the prediction looks correct, approve it directly — no edit needed. This sets FINAL.is_approved = True with FINAL.source_classification pointing at the AI Classification.

If something's wrong (wrong species, a missed animal, a bad bbox), fork the AI Classification into an editable USER Classification, fix it, then approve that instead. The original AI Classification is preserved untouched — forking never destroys it.

Either way, this is the moment the resource's classification becomes part of your project's accepted record. See Classification model for what's actually happening to the underlying rows.

6. Correct it later if needed

Found a mistake after approval? You don't need to re-open the case — log a FEEDBACK Classification against the approved row instead. It doesn't change the approval status; it's a permanent correction trail for quality control.

Verify it worked

The resource's classify view shows an approved (FINAL) classification with real species/observation data, and the Classification Project's dashboard resource counts reflect it as classified.

Next steps