The Lab
Where Half-Baked Ideas Come to Cook
Experiments, hypotheses, and things I'm tinkering with. Not everything works — and that's the point.
RAG pipeline for auto-classifying feature requests
Built a retrieval-augmented generation pipeline with K-means++ clustering that auto-classifies user stories by theme and urgency. Shipped in Backlog SaaS across 48 instances.
Can AI-generated outreach match human copywriters?
A/B tested Llama3-generated personalized outreach against manual copy for 50+ users. AI eliminated 100% manual effort while enabling real-time engagement tracking.
Redefining retention metrics for peer-to-peer apps
Splitwise loses 60% of users post-settlement because the product lifecycle ends at transactions. What if the North Star was group continuity, not settlement frequency?
VIP lead routing with latency < 5 min
High-value leads were getting lost due to slow routing. Prioritized a VIP alert system that reduced response latency by 95%, directly improving event-to-sales conversion.