Dawah Pal

Inspiration

Knowing an idea and explaining it clearly are different skills. Dawah Pal gives adults a place to rehearse an explanation out loud, with a listener who asks a focused question and feedback they can inspect. Parents and educators preparing conversations at home are one intended use case. We are exploring this as a contribution to at-home education, not a complete school or curriculum.

What it does

Choose a topic, read its lesson notes, and select one point to practise. You can speak with an AI listener in English or type an answer instead. Curious, Skeptical and Challenging listeners offer different styles of follow-up.

A voice session lasts up to three minutes. It can finish earlier when the listener considers the selected objective sufficiently explained or the learner asks to stop. The app then offers a separate review of the conversation. Feedback distinguishes points that came across, points needing clarification and points for which there is not enough evidence. You can inspect quotations from your own answer, view the referenced lesson material, retry one point and optionally save the outcome on your device.

The catalogue contains 732 pre-existing topic packets with lesson revisions, criteria, limits and source references. Search, subject groups and bounded result pages help learners choose an exercise without scrolling through the whole library.

How we built it

Codex was our development assistant for implementation, tests, debugging and preparation. The interface uses React, TypeScript and Vite. Browser Web Audio captures microphone input, draws signal-driven waveforms and schedules the returned PCM audio. A local Node.js server relays the conversation over WebSockets to Google Gemini Live (gemini-3.1-flash-live-preview). A separate Gemini Flash-Lite request (gemini-3.1-flash-lite) produces structured feedback.

The app sends the selected topic context, not all 732 topics, for a conversation. API credentials stay on the local server. The server constrains session duration and usage, checks topic revisions, and validates feedback source IDs and exact learner quotations. These checks improve traceability; they do not prove that every AI interpretation is correct.

Challenges and what we learned

The hardest part was making a real conversation behave reliably: interrupted audio must stop cleanly, an opening question must not give away the answer, and a useful exercise should not continue merely to fill a timer. We added bounded audio scheduling, neutral prompts, focused assessment and an explicit early-completion signal. We also separated a missing answer from an actual misconception, so an unasked optional point is not automatically counted as a mistake.

Validation and current limits

The current offline suite has 78 passing tests for audio lifecycle, local admission, topic selection, search, feedback validation, local state and completion behavior. Browser fixtures cover light/dark layouts and finished-call states. The founder has completed a real rehearsal and is testing the laptop flow again.

We do not claim that the tests establish model accuracy across every topic, learning effectiveness, validated market demand or a measured per-session cost. Feedback may be wrong and is not a scholar's verdict. This is a controlled adult-use prototype. It does not judge anyone's faith or sincerity.

Reuse, sources and AI disclosure

The 732-topic collection comes from the founder's existing Dawah Vox content project; it was not newly authored during the hackathon. Its source material includes YouTube lectures/transcripts, religious texts and scholarly references. Source IDs, URLs and roles remain with the topic packets. The phone-preview framework and device assets are reused infrastructure provided through the Codex prototype workflow.

Codex is the development tool. Gemini is the runtime provider for voice and feedback. Third-party frameworks, APIs and source material are credited in the repository's CREDITS-AND-REUSE.md and dependency manifests. Public access does not transfer ownership of third-party material. The repository preserves the actual development history rather than presenting reused work as newly created.

Try it and next steps

The GitHub README explains how to run the app on a laptop with your own Gemini API key. Our demonstration runs locally at 127.0.0.1:4180; this is not a hosted URL that remote judges can open. No phone installation is required for the demo.

Next, we want to test usefulness with adult learners and qualified educators, review disputed coaching, and measure complete session cost and response latency. Any proposed pilot budget or funding model is an estimate, not secured sponsorship or observed demand. We would expand access only after those checks.

Built With

• codex
• gemini
• node.js
• react
• typescript
• vite