Inspiration
- People with ADHD often struggle with task initiation, overwhelm, and context switching. Most AI tools just chat - they don’t help you actually start or finish anything. We built Sage to be a desk robot that doesn’t just talk, but executes, turning vague intentions into concrete progress with short, ADHD-friendly replies and zero overwhelm.
What it does
Sage is a beautiful, user-friendly AI productivity desk robot with full audio and visual control. It offers a calming frontend with voice input/output and four adaptive modes - Focus, Calm, Companion, and Plan.
Users can chat with Sage or create autonomous tasks: tell it a goal, and Sage plans step-by-step with Nemotron 3 Ultra, then executes each step using Nano and Super models on Nebius Token Factory to keep latency low and credits stretching further. It also includes a local usage ledger so users can track every model call, token, and latency metric independently.
How we built it
- Backend: FastAPI running on Render (Docker), powered by Nemotron models via Nebius Token Factory. Intelligent model routing selects Ultra for planning, Super for complex execution, and Nano for simple steps/greetings.
- Frontend: React + Vite hosted on Vercel, with voice input/output, visual feedback, and a calming UI design.
- Task Agent: Autonomous planner/executor that decomposes goals into numbered steps and runs them sequentially, with pause/resume capability.
- Usage Tracking: Append-only local JSONL ledger exposed via GET /api/usage, independent of the Nebius dashboard.
Challenges we ran into
- Balancing model tier selection to minimize Nebius credit usage while maintaining responsiveness.
- Keeping Sage’s replies short (1–2 sentences) to match ADHD-friendly communication without losing usefulness.
Accomplishments that we're proud of
- Built a fully autonomous task agent that plans with Ultra and executes with Nano/Super on real Nebius infrastructure.
- Integrated voice + visual controls into a calming, mode-adaptive interface.
- Local usage ledger gives users full transparency into token consumption and latency.
- Successfully deployed backend on Render and frontend on Vercel with a shareable live link.
What we learned
- Tiered model routing (Ultra/Super/Nano) dramatically reduces cost and latency for everyday tasks.
- ADHD-friendly UX requires brevity, calming visuals, and mode-based context switching - not just a generic chatbox.
What's next for Sage - A Productivity Desk Robot
- Add recurring task scheduling and reminders.
- Integrate with calendar and todo apps (Google Calendar, Todoist, Notion).
- Expand voice controls with wake-word and hands-free operation.
- Offer premium tiers with dedicated Nebius Serverless Endpoints for guaranteed uptime.
- Add team/shared workspaces for collaborative task execution.
Built With
- fastapi
- github
- nemotron
- nvidia
- openai-sdk
- python
- react
- tokenfactory
- typescript
- vite
Log in or sign up for Devpost to join the conversation.