Inspiration
Creators don’t fail because their content is bad - they fail because the algorithm is opaque. Small factors like title structure, tags, and publish timing can drastically change performance, yet most creators rely on guesswork.
That gap inspired Arceus - a system that brings data-driven clarity to virality, helping creators understand how the algorithm sees their content before they publish.
What it does
Arceus is a virality intelligence tool for YouTube creators.
Given a planned video (title, tags, description, category, timing), it:
- Predicts where the video will rank within its cohort
- Finds “viral twins” - similar videos that already performed well
- Suggests specific, model-validated improvements (e.g., timing, tags, title tweaks)
- Re-scores the improved version to show measurable gains
It turns uncertainty into action:
“Here’s exactly how to make your video perform better.”
How we built it
We built Arceus as a three-stage ML pipeline:
Predictor (Performance Model)
- LightGBM regression trained on 178K trending videos across 11 countries
- Features:
- Title structure (length, caps, punctuation)
- Timing (hour, day)
- Context (country, category)
- Content embeddings (TF-IDF + SVD)
- Output: percentile rank, not raw views
Viral Twins (Retrieval System)
- Multilingual sentence-transformer embeddings
- Matches input with similar high-performing videos
- Ranked by semantic similarity + actual success
Optimiser (Decision Engine)
- Greedy search across:
- publish time
- publish day
- tags
- title structure
- Keeps only edits that increase predicted performance
- Produces a final improved version with measurable lift
- Greedy search across:
Product Stack
- FastAPI backend (model serving)
- React + Vite frontend (interactive dashboard + globe view)
- Deployed via Docker on Hugging Face Spaces
Challenges we ran into
- Data leakage: Same videos across multiple countries - required group-based splitting
- Noisy descriptions: Spam, links, timestamps - needed aggressive cleaning
- Extreme skew in views: Solved using log scaling + percentile targets
- Speed vs accuracy trade-off: Transformers are powerful but slow - hybrid architecture
- Generic recommendations failing: Fixed by making all suggestions model-validated
Accomplishments that we're proud of
- Built a complete end-to-end ML product, not just a model
- Achieved:
- AUC: 0.88
- Spearman: 0.78
- Improved baseline by +0.09 to +0.22 AUC across all countries
- Created a system where:
- Predictions are interpretable
- Suggestions are actionable and measurable
- Successfully deployed a live, production-ready app
What we learned
- Content understanding drives most of the predictive power
- Data quality matters more than model complexity
- Users don’t just want predictions - they want clear actions
- Simpler models can outperform complex ones with the right features
- Building real tools requires balancing accuracy, speed, and usability
What's next for Arceus: Virality Intelligence
- AI-powered title rewriting using fine-tuned models
- Channel-level personalization (learning creator-specific patterns)
- Improved multilingual understanding and cleaning
- Real-time trend integration
- Direct platform integrations for seamless creator workflows
Vision
Arceus isn’t just a tool - it’s a shift in how creators approach content.
