Inspiration
Most habit-tracking apps store data in the cloud, offer limited real control to the user, and focus mainly on passive logging instead of actively helping people stick to their goals. In the Nebius x NVIDIA Global AI Hackathon (Personal AI Track), we wanted to build a true Personal AI Steward — an always-available, privacy-first assistant that learns from the user’s own behavior and helps them improve over time. The core idea is that habits form loops. The AI doesn’t just record what happened; it proposes actions, verifies them, corrects when needed, and automatically learns constraints from repeated failures so the user can build better, more sustainable habits.
What it does
HabitLoop Steward is a privacy-first personal AI habit tracking assistant that works both online and offline:Quick habit logging via CLI or Web UI
All data stored locally in SQLite (no telemetry, no cloud storage of habit data)
Code-based calculation of streaks and adherence rates
AI coaching powered by NVIDIA Nemotron (via Nebius Token Factory) for daily nudges, plan adjustments, and insights
Graph/Loop Runtime: Propose → Verify → Correct workflow with a separate verifier and RED/YELLOW/GREEN safety gates
Constraint Learning: Automatically detects low-adherence patterns and creates helpful rules
Weekly reviews combining deterministic stats + optional AI synthesis
Serverless jobs for nightly reviews and morning briefs (deployable to Nebius)
Clean FastAPI Web UI with a dedicated Privacy panel
Everything works gracefully in fully offline mode when no API key is provided.
How we built it
We built the project in pure Python 3.9+ with a clean modular architecture:SQLite for local, private storage
Typer for a rich CLI
FastAPI for the web interface
Nebius Token Factory + NVIDIA Nemotron for optional AI capabilities
Custom Graph/Loop Runtime implementing Propose → Verify → Correct with role separation
Domain skills (nudge, plan-adjust, stats)
Constraint learning and weekly review engines
Serverless job handlers with local fallback
The system is designed so core tracking features never depend on the LLM, while AI features enhance the experience when available.
Challenges we ran into
Designing a reliable Propose → Verify → Correct loop with a truly separate verifier role
Making the system degrade gracefully to full offline mode without any API key
Automatically detecting meaningful patterns from sparse habit data to create useful constraints
Balancing deterministic code-based calculations with optional AI insights
Keeping the entire experience privacy-first while still offering powerful AI coaching
Accomplishments that we're proud of
A complete, working MVP with both CLI and Web UI
Full privacy-first design (local SQLite, zero telemetry)
Robust Graph/Loop Runtime with safety gates
Automatic constraint learning from real usage patterns
Serverless job support (nightly review + morning brief)
Comprehensive offline-capable test suite
Clean architecture that clearly separates deterministic logic from AI components
What we learned
How to design AI systems that remain useful even when the model is unavailable
The value of separating “proposer” and “verifier” roles for safer AI behavior
Practical techniques for learning simple constraints from user behavior data
Best practices for building privacy-first personal AI tools
How to combine Nebius Token Factory and NVIDIA Nemotron effectively in a real product
What's next for HabitLoop Steward
Mobile-friendly progressive web app
Richer visualization of streaks, constraints, and weekly trends
Multi-user / multi-device sync with end-to-end encryption
More sophisticated constraint learning and habit recommendation engine
Deeper integration with Nebius serverless for always-on steward capabilities
Optional voice interface and more personalized coaching styles
Log in or sign up for Devpost to join the conversation.