Predict which of two video edits will hook viewers before posting.
Generate multiple voiceover reads for a video and score which one holds attention longest.
Compare engagement curves across five brain systems for different content versions.
Prototype a two-backend architecture splitting a no-code frontend from a Python ML backend.
| sxnnywu/reeled-in | a-shojaei/constructdrawingai | alex72-py/aria-termux | |
|---|---|---|---|
| Stars | 20 | 20 | 20 |
| Language | Python | Python | Python |
| Setup difficulty | hard | moderate | moderate |
| Complexity | 4/5 | 4/5 | 2/5 |
| Audience | developer | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires GPU access on Modal, API keys for ElevenLabs and Auth0, and a MongoDB Atlas database.
Reeled In is a hackathon project built for Hack the 6ix 2026 in Toronto that offers neural A or B testing for short-form video. Instead of guessing which version of a video edit will perform better, or paying for a focus group, a creator uploads two or more versions of a Reel or TikTok, or provides a single video with a voiceover script, and the tool predicts how a viewer's brain will respond to each version before anything is posted. It uses Meta's open source TRIBE v2 brain model, breaking the predicted response into five systems: visual, auditory, language, motion, and a default mode network, then measures engagement over time to pick a winner. The README is careful to state this predicts brain activation rather than sales or virality, describing it as a fast directional signal rather than ground truth. The minimum viable product supports two test types. The first lets a creator upload two or more finished video versions for the tool to score and compare directly, without any automatic editing. The second, called Voice A or B, takes one base video and a script, generates several different voice readings using ElevenLabs, overlays each onto the video, and scores the results. For either type, the tool computes peak engagement, sustained engagement, and how well a version holds attention through to the end, and a user picks which of these goals matters most to determine the winner. The system is split into two backends by design. A Base44 frontend handles uploads, login through Auth0, and the results screen, while a Python backend built with FastAPI and deployed on Modal holds all the secrets and heavy processing, calling a Modal GPU to run the TRIBE v2 scoring, ElevenLabs for voice generation, a service called Backboard for suggestions and memory, and MongoDB Atlas to store tests, versions, and scores. The reasoning given is that the frontend platform cannot run heavy machine learning workloads or reliably reach the database, so all of that logic lives in the Python service instead. Setup for the scoring engine happens on Modal, following a separate setup guide, and requires being added to the team's Modal workspace. The README notes that TRIBE v2 is licensed for non commercial use only, so the project is treated as a free beta rather than a paid product.
A hackathon tool that predicts how viewers' brains respond to different short-form video edits before you post them.
Mainly Python. The stack also includes Python, FastAPI, Modal.
Not stated for the project's own code, the TRIBE v2 model it depends on is licensed for non-commercial use only.
Setup difficulty is rated hard, with roughly 1h+ to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.