Get end-to-end BBQS workflow example running

[CodyCBakerPhD]

Parent issue for each individual step

There are two independent tracks for this, each of which can be tested and tuned:

• SLEAP in the cloud

  • The official project from the SLEAP team
  • Dev app hosted on https://app.sleap.ai/dev/ (v0.1.0 as of 6/17/26)
  • I suppose the source code would be on https://github.com/talmolab/sleap-app, but this seems to be private ATM
  • More general, allows users to bring their own videos and compute resources for training. Allows users to define their own skeletons and video encodings.
  • Classic SLEAP workflow, end-to-end (vertically scaling, models trained per project)
  • No EMBER/DANDI/Remote support yet available but planned

• Pozu (working name until confirmation from Zooniverse)

  • Prototype project by me
  • Temporarily hosted on https://codycbakerphd.github.io/pozu/
  • Public source code for frontend: https://github.com/CodyCBakerPhD/pozu
  • Very specific intention of rapid, collaborative labeling. Speculating on some training integration capability for specific datasets or collections. Trying to make it as fast as possible to go through labeling process, which may include multiple modes (split screen, focus)
  • Strong emphasis on standardized workflow; only one (or very few) skeletons allowed, all videos transcoded to the same resolution (scale discussion pending)
  • Break from the standard SLEAP/DLC/LP workflow; breaks up the end-to-end parts into distinct steps, starting with consolidation of large number of highly standardized training frames, followed by high-volume training of the ML models in attempts to make generalized foundation models (horizontally scaling)
  • No support for local videos; all videos must be uploaded and run through EMBER/DANDI. Interactions within the app are automatically contributed directly back to the archive through the backend proxy server

In both cases, we would like to attempt to use the Pennsieve workflow @rhingo had previously prototyped to run the ML training and inference, though that is currently conditional on some key fixes to the platform and could be done through other means in the intermediary or long-term