Personal ProjectFeatured

Podcast-to-Pack: AI-Powered Podcast Analysis Platform (Award Winner)

Builder & Product Architect|

3 months

6h → 90min (85% reduction)

Research Time

Strongest Business Case (Agentic AI PM cohort, business model + tech vs. 40+ competitors)

Award

FastAPI + Pydantic AI

Architecture

4.8/5 avg rating (n=28 beta users)

User Feedback

The Challenge

Podcast listeners struggle to extract actionable insights from hours of content. Manual note-taking is slow, inconsistent, and misses connections. Needed a system that could analyze podcasts, extract key insights, identify themes, and generate structured deliverables — all while handling messy real-world audio and maintaining accuracy.

The Approach

Built a FastAPI + Pydantic AI system with Deepgram transcription, Claude Sonnet 4 for analysis, and Stripe for payments. Architecture: (1) Audio ingestion with format normalization, (2) Deepgram transcription with speaker diarization, (3) Multi-stage analysis pipeline (themes → insights → summaries → deliverables), (4) Structured output with Pydantic validation. Added circuit breaker pattern for API resilience and structured logging for debugging.

Key Learnings

Circuit breaker pattern prevents cascade failures with flaky LLM APIs
Speaker diarization quality matters — Deepgram's model made analysis 40% more accurate
Structured outputs with Pydantic catch hallucinations before they reach users
FAILURE: Initial prompt engineering for insight extraction was too generic — had to add domain-specific templates for podcast niches (tech, business, education)