Inspiration
I often save research papers, earnings reports, policy documents, and long PDFs with the intention of reading them later, but "later" rarely comes. While podcasts are an easy way to learn on the go, most AI-generated summaries don't clearly show where their conclusions come from. I wanted to build something that was as convenient as listening to a podcast while remaining grounded in the original document.
What it does
InsightCast transforms long documents into evidence-backed, multi-speaker podcast briefings. Users upload a PDF, and the system extracts the content, identifies key facts, and generates a natural conversation between three AI personas:
- Host – guides the discussion.
- Analyst – explains important findings.
- Skeptic – questions assumptions and highlights limitations using evidence from the source.
The final experience includes a podcast player with chapter navigation, transcript highlighting, and replayable briefings, making long documents much easier to consume.
How we built it
InsightCast is built as a structured multi-agent pipeline rather than a single AI prompt.
Document
↓
Extraction
↓
Research
↓
Planning
↓
Writing
↓
Review
↓
Multi-Voice Podcast
The document is parsed using pdfplumber, after which specialized agents research the content, organize key ideas, write the conversation, and review the final script before audio generation. Stable fact references are carried through the pipeline so important claims remain traceable to the source. The frontend is built with React, while the backend orchestrates the AI workflow and generates the final podcast.
Challenges we ran into
The biggest challenge was building a reliable system while working with limited hardware and API constraints. Smaller local models occasionally generated unsupported claims, while hosted models could hit token and rate limits.
To improve reliability, I split responsibilities across specialized agents, added structured output validation, implemented retry and response repair logic, passed stable fact references between every stage, and limited reviewer rewrites to prevent endless loops. I also built deterministic demo fixtures so the complete workflow could be tested consistently without depending on live model responses.
Accomplishments that we're proud of
- Built a complete end-to-end pipeline that converts documents into multi-speaker podcast briefings.
- Designed a grounded multi-agent architecture that keeps AI-generated conversations tied to the source document.
- Created distinct AI personas that make technical content more engaging without sacrificing accuracy.
- Implemented a podcast player with transcript highlighting, chapter navigation, and replayable briefings.
- Built a deterministic testing workflow that makes demonstrations and evaluation reliable.
What we learned
One of the biggest lessons was that trustworthy AI isn't just about writing better prompts. Grounding has to be designed into the entire system through structured data, validation, review, and traceability. Splitting the workflow into specialized agents and enforcing fact references throughout the pipeline produced much more reliable results than relying on a single large prompt.
What's next for InsightCast
The next step is expanding beyond PDFs to support URLs, Word documents, research papers, and YouTube transcripts.
I'm also planning to add persistent agent memory so users can save briefings, revisit previous documents, compare new reports with earlier ones, and ask follow-up questions across multiple sessions.
Longer term, InsightCast will generate multiple outputs from the same source—podcast briefings, executive summaries, newsletters, presentation slides, and short social media clips—while preserving source citations and traceability across every format.
Log in or sign up for Devpost to join the conversation.