new-luna-creative-project

InspirationInspiration

Our inspiration stemmed directly from Dario Amodei’s vision in Machines of Loving Grace and the core dilemma of Track 5 (Creative Flourishing): As AI handles routine work, questions about human purpose become urgent. We noticed a massive flaw in current AI writing tools—they quietly replace human creators by generating the content for them.

We wanted to build an AI tool that holds up a mirror, not a pen. We were deeply inspired by rich Indian oral storytelling traditions like Baul, Harikatha, and Dastangoi and wanted to create a platform that bridges the gap between written text, musical performance, and language preservation.

🛠️ How we built it Katha is built as a lightweight, fast, and responsive web application:

Backend: We used Python and Flask to create a robust and minimal server to handle user requests and routing.

AI & Mentorship Engine: We engineered a strict prompt pipeline that forces the LLM to output structured JSON analyzing six dimensions of writing (sentence rhythm, emotional vocabulary, sensory engagement, cultural depth, perspective, and metaphor density). It is strictly governed by a zero-generation rule—it only analyzes and challenges.

Frontend: We used vanilla JavaScript and CSS, integrating Chart.js via CDN to instantly render beautiful, dynamic data visualizations (like our Sensory Radar and Rhythm Doughnut charts).

Database: We implemented a file-based JSON archive (stories.json) for quick MVP deployment to act as our living community library.

🚧 Challenges we faced

Quantifying Art: Our biggest technical hurdle was getting the AI to reliably translate highly subjective narrative qualities (like "cultural depth" or "sensory engagement") into strict, quantifiable JSON metrics without breaking the frontend formatting.

Performance Context: It was challenging to make the AI infer musical and stage directions (pacing, volume, pauses) purely from written text, but we solved this by grounding the AI in specific folk performance traditions.

Prompt Guardrails: Ensuring the AI never wrote a single line of creative content for the user required extensive prompt engineering and strict negative constraints.