DayOne
2025archivedUpload one training video, get a Duolingo-style micro-learning course with interactive challenges and an AI mentor that cites the video back to you.
Story
A manufacturer has decades of know-how locked inside the people about to retire. Recording a training video is the easy part. Turning it into something a new hire can actually learn from is where the work dies.
DayOne ingests a single video and uses Gemini 2.0 Flash to extract the procedural knowledge inside it — then generates 15 different challenge types so the learner *practices* instead of just watching. Drag-and-drop step ordering. Dial calibration. 3D-style interactions. Timed recall. Quiz checkpoints. An AI mentor sits alongside the course with timestamped citations back to the source video so every answer can be verified against the real demonstration.
Offline-first by design: IndexedDB caches all challenges and the whole experience runs on a shop-floor tablet with weak Wi-Fi.
What I built
- 1Built the multimodal AI pipeline generating 15 interactive challenge types with strict Zod-validated schemas
- 2Designed challenge types including drag-order, dial calibration, 3D-style interactions, timed recall, and quiz checkpoints
- 3Built the AI mentor with timestamp-grounded citations back to the original video
- 4Implemented client-side video frame extraction with accurate cross-browser timestamp seeking
- 5Designed the offline-first architecture: IndexedDB caches all challenges, cloud sync runs opportunistically