Inspiration
ATLAS began as a modern revival of CLIPPY — not as a joke, but as an exploration of what a truly useful desktop assistant could look like today. The goal was to build an assistant that lives on the user’s screen, understands real context like files and activity, and improves over time instead of interrupting.
What it does
ATLAS is a multi-mode desktop assistant that reacts to what the user gives it. Users can drag and drop files directly onto the assistant, and ATLAS processes them intelligently depending on the selected mode.
It can:
Analyze documents and generate structured research summaries
Parse bank statements and produce spending insights and visualizations
Handle domain-specific workflows like lost-and-found matching
Collect focus and activity telemetry to support long-term personalization
ATLAS is designed to progressively gain more capability as it “lives” with a user, learning habits and preferences rather than relying on static prompts.
How I built it
ATLAS was built entirely solo in Python. It uses a PySide6 desktop interface with a persistent sprite assistant, background worker threads and subprocesses to keep the UI responsive, and modular task modes that can be extended independently.
The system combines local processing, structured data pipelines, and LLM-assisted reasoning where appropriate. An active-learning subsystem collects optional telemetry (focus, breaks, app usage, noise levels) and stores it for future personalization.
Challenges I ran into
The biggest challenges were handling messy real-world data formats, especially bank exports and documents that don’t follow clean schemas. Managing concurrency without freezing the UI required careful separation between UI threads, worker threads, and subprocesses. Designing a system that stayed modular while still feeling cohesive was also non-trivial. Debugging edge cases across multiple modes took significant time.
Accomplishments I’m proud of
ATLAS is a fully working desktop assistant that supports drag-and-drop file input and runs multiple intelligent workflows inside a single interface. It robustly processes real user data, not just ideal test cases. The architecture is modular, extensible, and ready for future capabilities like deeper personalization and proactive assistance — all built by one person.
What I learned
Real data is always messier than expected. Desktop applications require careful concurrency design when combined with AI workloads. UX decisions matter just as much as model quality. Most importantly, building a flexible foundation early makes it much easier to grow a system over time.
What’s next for ATLAS
Next steps include deeper user memory, preference modeling using the active-learning data, smarter non-intrusive suggestions, and expanding ATLAS into a true personal operating assistant that adapts to how each individual works.
Log in or sign up for Devpost to join the conversation.