How I Built a Lightweight Edge AI Pipeline for Podcasts

Description and use

A lightweight edge AI pipeline for podcasts that generates show notes and titles using LLMs, speeding up post‑production.

Problem

Post‑recording tasks—creating show notes, titles, and timestamps—drained indie podcasters and were hard to automate with existing AI tools.

Outcome

The pipeline produced concise, structured show notes and compelling titles with minimal latency, freeing the creator from tedious manual editing.

Workflow

• 1. Record podcast and export transcript
• 2. Send transcript to Cloudflare Worker
• 3. Worker calls @cf/meta/llama-3-8b-instruct with constrained system prompt to produce markdown show notes
• 4. Worker returns JSON with episode summary, key takeaways, timestamps, resources
• 5. Separate title generator builds prompt with tone and count, calls LLM, returns numbered list of titles
• 6. Podcaster reviews and publishes metadata

Evidence

I looked at existing AI tools to help with this, but they were either generic text spinners that hallucinated timestamps, or massive enterprise suites priced for media companies, not solo developers.

Source

Here is the core of the show notes generation pipeline. I pass the raw transcript and use a highly constrained system prompt to force the model into a specific markdown structure:

Source

Breaking the pipeline into smaller, constrained prompts worked much better than my initial attempt to do everything in one giant prompt.

Source