AI/ML · Personal Project
Post-Release Optimizer
AI chatbot that turns 188K track dataset into personalized release strategies for artists
The Challenge
Independent artists and small labels lack data-driven tools for post-release strategy. After dropping a song, they have no systematic way to analyze comparable releases, identify playlist targets, or time promotional pushes — missing momentum during the critical first-week window.
My Role & Approach
Made the deliberate product decision to build a conversational interface rather than a dashboard. Artists think in questions ('What playlists should I target for my indie-pop track?'), not in charts. This insight drove the entire architecture toward an NLP-powered chatbot.
Product Thinking
Problem Framing
The initial idea was a Spotify analytics dashboard. Through conversations with independent artists, I learned they don't think in charts — they think in questions: 'What playlists should I target?' 'When should I push on Instagram?' This insight pivoted the entire product from dashboard to conversational interface.
Key Tradeoffs
Conversational AI vs. traditional dashboard: dashboards are easier to build but don't match how artists think. Chose the harder path of NLP-powered chat, accepting that some query types would be less precise than a filtered table. The friction reduction was worth the accuracy tradeoff.
What We Didn't Build
Scoped out Spotify API integration for real-time streaming data, social media auto-posting, and A/B testing for release strategies. The core insight was that artists need strategy recommendations, not more data — so I focused exclusively on the recommendation engine.
Solution & Execution
AI chatbot that ingests 188,000+ tracks and generates personalized post-release strategies. NLP entity extraction for genre, mood, and tempo matching. Recommendation engine for playlist targeting, social media timing, audience segments, and promotional spend allocation. Streamlit interface for rapid iteration.
Tech: Python, NLP, LLMs, Pandas, Streamlit
Impact
- 188,000+ tracks analyzed across multiple platforms and genres
- 15+ query types supported for personalized release strategies
- Natural language interface reduced friction for non-technical users
Challenges & Pivots
Challenge: Music industry data was extremely messy — duplicate tracks, inconsistent metadata, missing fields
Resolution: Built a deduplication pipeline and normalization layer. Significant effort but essential — garbage in, garbage out applies especially to recommendation systems.
What I Learned
- Conversational interfaces beat dashboards when users think in questions, not metrics.
- Domain-specific chatbots benefit enormously from structured data backends rather than relying purely on LLM knowledge.
Next case study
LitLens — Literature Intelligence Platform