How My Bluesky Digest (V2) MVP Failed

My previous post ended with a promise: I was going to rebuild the ranking from scratch. Train a proper ML model. Use BM25 and embeddings as features, not as the final answer. And use LLMs to generate