Nudge Ai

Inspiration

We realized that traditional to-do lists are passive "graveyards" where tasks go to die because they rely on the user to check them. We wanted to flip the script by building an "active" agent that behaves like a relentless executive assistant, hunting the user down to ensure work actually gets done.

What it does

NudgeAI is an autonomous agent that parses natural language commands into actionable workflows, automatically scheduling and executing tasks like sending emails or booking meetings. Unlike standard web apps, it runs as a persistent background process, triggering real-world interruptions and actions on your OS even if the browser dashboard is closed.

How we built it

We architected a full-stack solution using FastAPI for the backend logic and React for the frontend, connecting them via a RESTful API. The system uses OpenAI to intelligently route user intent into structured JSON, while APScheduler and Plyer handle the asynchronous execution of tasks and native desktop notifications.

Challenges we ran into

Our biggest technical hurdle was preventing the AI from hallucinating task details; we solved this by implementing a strict "Router Prompt" that forces the model to output structured JSON data. We also had to overcome significant CORS restrictions and Windows security permissions to allow the web frontend to trigger local OS-level alerts.

Accomplishments that we're proud of

We are proud of achieving "Browser Independence," proving that our agent continues to execute complex tasks and send real emails in the background long after the user closes the tab. We also successfully transformed a simple timer into a "Tool Router" that dynamically switches between creating reminders and executing real API calls based on context.

What we learned

We discovered that robust "prompt engineering" is just as critical as backend code; tweaking the system prompt solved more logic bugs than rewriting Python functions. We also learned the immense value of "latency hiding"—giving users immediate visual feedback while the heavy AI processing happens asynchronously in the background.

What's next for Untitled

We plan to integrate deep two-way connections with Slack and Google Calendar to allow the agent to negotiate meeting times autonomously. We also aim to build an "Escalation Protocol" where the agent increases the severity of notifications—from gentle nudges to public alerts—if high-priority tasks are repeatedly ignored.

Built With

• ai
• axios
• css3
• engine:
• fastapi
• frontend:-react.js-(vite)
• gpt-3.5
• lucide-react.-backend:-python-3.10+
• openai
• turbo
• uvicorn.