The problem
TwoSetAI runs a podcast with 100+ interviews with AI founders and leaders (Bolt, Hightouch, Merge, Yutori, tldraw, and others). The interviews contain patterns that repeat across founders, but extracting those patterns by hand is slow and summarizing each transcript individually loses the cross-interview signal. We want an open-source system that ingests the corpus, builds a searchable knowledge base, and answers questions like 'how do early AI founders find their first customer?' with grounded citations. Needs a benchmark question set and measurable precision/recall.
What success looks like
Retrieval precision@5 above 70% on a 20-30 question benchmark. Open-source repo. Written report.
Data access
100+ cleaned transcripts, each tagged with company and founder. Provided under academic use agreement.
Constraints
No PII. Open-source license required (MIT or Apache 2.0).
Reward
Featured in a TwoSetAI workshop + GitHub stars