{ "project": "Bedtime Story Machine", "track": "Thousand Token Wood", "builder": "ShebMichel", "tools_used": [ "Kiro CLI (AI agent)", "Python", "Gradio", "Modal", "ffmpeg" ], "models": { "story_generation": "nvidia/Nemotron-Mini-4B-Instruct (4B params)", "illustration": "black-forest-labs/FLUX.1-schnell (12B params)", "total_params": "~16B" }, "infrastructure": { "inference": "Modal (Nemotron on T4 GPU)", "image_gen": "HF Inference API", "hosting": "Hugging Face Spaces", "frontend": "Gradio with custom CSS" }, "build_steps": [ { "step": 1, "action": "Designed app architecture - chose Nemotron-4B for text, FLUX.1-schnell for images" }, { "step": 2, "action": "Built Gradio app with storybook UI (custom CSS, alternating scene layout)" }, { "step": 3, "action": "Implemented story generation with structured JSON output + regex fallback" }, { "step": 4, "action": "Integrated FLUX.1-schnell for watercolor-style illustrations" }, { "step": 5, "action": "Added French language support via language dropdown" }, { "step": 6, "action": "Set up Modal deployment for Nemotron-4B with vLLM" }, { "step": 7, "action": "Created demo video with ffmpeg subtitle burning" }, { "step": 8, "action": "Deployed to HF Spaces under build-small-hackathon org" }, { "step": 9, "action": "Published blog post and agent trace" } ], "challenges": [ "Nemotron-4B not available on any free inference provider - solved with Modal deployment", "HF token needed fine-grained Inference API permissions", "JSON parsing reliability from small models - added regex extraction fallback", "Consistent illustration style - solved with prompt engineering suffix" ], "badges_targeted": [ "\ud83d\udfe9 NVIDIA Nemotron Quest (Nemotron-4B model)", "\ud83d\udc1c Tiny Titan (\u22644B story model)", "\ud83d\udfe2 Modal Awards (Modal deployment)", "\ud83d\udcd3 Field Notes (blog post)", "\ud83d\udce1 Sharing is Caring (this trace)" ] }