Case study

How Podcasters started.

The product now generates structured dialogues, surprising facts, and fully playable episodes — all inside the browser.

What if anyone could turn a topic into a podcast instantly?

The question that started it

Product vision

Knowledge that feels like a conversation

Make knowledge accessible through guided conversations that feel like a personal tutor. Podcasters condenses complex topics into human, engaging back-and-forth dialogue — the format people already learn best from.

The journey

From question to product

Discovery

Problem discovery

People learn best through conversation, but most topics don't have a ready-made podcast. Learners wanted short, structured summaries without hunting across platforms to find them.

Prototype

Prototype & iteration

Early mockups tested whether users preferred structured explainers or fast-paced fact streams. The two-mode system — "Knowledge" versus "Interesting Facts" — became the product's signature.

Validation

Technical validation

Web Speech API playback, multi-voice selection, dynamic fact injection, and structured prompting across OpenAI GPT and Claude — with output-schema validation to keep responses consistent across models — proved that fully client-side podcast generation was feasible and fast, with topic-to-audio time under 60 seconds.

My role

Solo engineer — concept to production

I designed and built Podcasters end-to-end: the LLM pipeline, the prompt and schema-validation layer, and a fully client-side product shipped with zero backend infrastructure.

System architecture

  • Designed a zero-install, fully client-side architecture — no servers, no build step — so generation runs entirely in the browser.
  • Chained multi-step LLM outputs into a synthesis pipeline, mapping structured dialogue turns to per-speaker voice profiles via the Web Speech API.
  • Built local, in-browser API-key storage to keep credentials off any server, prioritizing a privacy-first design.

LLM integration & prompt engineering

  • Integrated multiple LLM providers (OpenAI GPT and Claude) with output-schema validation to normalize structured dialogue across models.
  • Engineered and iterated on prompts to hold topic-to-audio latency under 60 seconds while optimizing token usage.
  • Built a two-mode generation pipeline ("Knowledge" vs. "Interesting Facts") tuned from feedback across 120+ users.

Technical stack

  • Client-side JavaScript for speed, portability, and zero-install distribution.
  • Dynamic fact-injection pipeline combining DuckDuckGo retrieval with LLM post-processing for grounded, surprising facts.
  • Web Speech API playback with per-speaker voice assignment and transcript export.

Future roadmap

  • Add server-side TTS for higher-fidelity, distinct narrator voices beyond system voices.
  • Ship a shareable "auto-episode link" with server-rendered playback.
  • Build a mobile-friendly progressive web app with offline caching.
  • Support playlist generation for chained, multi-topic learning tracks.
Try it yourself

Hear what it sounds like

Pick any topic and have a two-voice episode ready to play in under a minute.

Create an episode