Founders Sage (replica)

A richer reimagining of Founders Notes' Sage — David Senra's $100/mo chatbot over his Founders Podcast notes. The original is a vanilla vector-RAG with shallow context; my version is an agentic engine that browses transcripts, raw source books, and Senra's own notes to actually answer the question. Built as an experiment, never pitched.

Context
Personal experiment — a better knowledge engine over the Founders Podcast corpus.
Role
Designed and built end to end, solo.

Bill of materials

  • Python
  • Vector + agentic retrieval
  • LLM tool use
  • Anna's Archive scraping

Founders Notes sells a product called Sage for $100/month — a chatbot over David Senra’s notes from his Founders Podcast (deep biographical podcasts on famous founders, drawn from books). It’s useful, but it’s a thin product: a vanilla vector-RAG that depends entirely on your query being phrased richly enough for the embedding lookup to land in the right place. Ask a slightly oblique question and it misses.

I built a replica with a much deeper knowledge engine.

What changed

  • Agentic retrieval, not single-shot RAG. The agent can browse the corpus, pivot, follow threads across episodes, and pull the right context rather than depending on one embedding match.
  • Three-layer corpus. Podcast transcripts, Senra’s own notes, and the raw source books themselves (scraped from Anna’s Archive). Sage only has Senra’s notes; mine had the underlying material those notes summarised, which let the agent ground claims and surface specifics Sage couldn’t.
  • Use-case-shaped responses. Because the agent can decide what to retrieve mid-turn, queries like “what did three different founders do when their cofounder left” actually compose across episodes instead of returning the closest single chunk.

Outcome

I considered pitching it directly to Senra and never did. It stays in the experiments drawer — a working prototype that demonstrates how much headroom there is above naïve RAG when the corpus deserves a real agent.