DuckPod

DuckPod is a human-first AI live podcast agent where users generate AI episodes and audit them with trust rubric. It scores reliability, flags hallucinations, and keeps humans in control of AI truth.

Comment

Inspiration

AI-generated content is fluent, fast, and persuasive — but often unchecked. We noticed that most AI tools optimize for output speed, not trust. DuckPod was inspired by a simple question: What if AI had to earn credibility instead of assuming it?

What it does

DuckPod is a human-first AI podcast Agent. It generates short AI-hosted podcast episodes and then requires a structured human audit. After listening to an episode, users evaluate it using a trust rubric: Grounding (facts vs vibes) Consistency Transparency Manipulation resistance Usefulness

The system computes a Trust Score (0–100) and flags potential hallucinations. AI speaks. Humans decide if it deserves trust.

How we built it

  • Frontend built using Lovable with a structured multi-page flow

  • Gemini API for podcast script generation and claim extraction

  • NodeJS and Supabase for session storage, reflections, and trust score persistence

  • Deterministic scoring logic for Trust Score and hallucination detection

  • Human-in-the-loop evaluation layer to enforce accountability

We designed it as a state-driven flow: Setup → Studio → Reflection → Trust Report.

Challenges we ran into

  • Making AI outputs structured enough for auditing

  • Designing a scoring system that felt fair but decisive

  • Preventing AI from “judging itself”

  • Balancing simplicity with meaningful evaluation

  • Ensuring the Trust Score felt transparent, not arbitrary

Accomplishments that we're proud of

  • Built a full human-audit layer over AI-generated content

  • Created a weighted Trust Score that prioritizes grounding

  • Implemented deterministic hallucination detection logic

  • Delivered a clean audit-style UI that reinforces human authority

  • Positioned AI as a tool, not a decision-maker

What we learned

  • AI fluency can mask uncertainty

  • Trust requires structure, not vibes

  • Human-in-the-loop design is powerful

  • Accountability mechanisms increase confidence

  • Clear scoring systems reduce ambiguity

What's next for DuckPod

  • Citation-required mode with live source validation

  • Adversarial questioning mode

  • Multi-model comparison (Gemini vs others)

  • Confidence gap detection (AI confidence vs human trust)

  • Shareable Trust Cards for AI literacy education

Built With

Share this project:

Updates