Spoilr
Thesis
A fantasy-sports-style prediction game for scripted television. Players call outcomes before each episode airs; the platform generates the questions, scores the answers, and runs the leagues.
The bet: episode discussion is already a game — Spoilr gives it rules, standings, and stakes.
Under the hood
Python service core on PostgreSQL. LangGraph pipelines ingest episodes, extract narrative threads with Anthropic models, and emit graded prediction questions. Scoring is deterministic; generation is reviewed before publish.
Decisions
- 01
Question generation is a graded pipeline, not one prompt.
Each stage is testable and each failure is attributable — worth the extra moving parts.
- 02
Postgres over a vector store for episode memory.
Relational queries answer 90% of retrieval needs; one fewer system to operate.