HypeHound

HypeHound polls a Resend-managed mailbox, strips newsletter HTML to plain text, then classifies each ~600-word chunk with a local LLM — one structured PydanticAI call per chunk, output validated against a Pydantic schema — and stores the extracted items in SQLite. A second pass groups items by category and writes a clean digest; a third turns the digest into a narrated dual-host script, synthesised locally via Voicebox (Qwen3-TTS). The mash-up of all newsletters are condensed into a bitesize digest and published here automatically.