Intention.ai

Inspiration

Studies found that workers globally can lose over 500 hours annually due to disruptions and lack of focus. We want to change that. Given the role LLMs like Google Gemini have played in the productivity and workflow improvement space, the nature of this existing space and our interest in computer vision enabled us to tackle a problem space we felt close to in a way that felt both relevant/adoptable and personally interesting.

What it does

Intention.ai allows hard-working, high-achieving students to be intentional about their active work time by providing suggestions to improve efficiency based on their perceived level of focus. Using a computer vision model to identify when a studying user is tired, focused, or stressed, the browser extension tracks focus patterns and provides personalized recommendations to help the user structure their study sessions more effectively. For example, if the user seemed to be tired for a prolonged amount of time, the system may suggest taking a short break. Users gain further insights through the intention.ai dashboard, which accumulates data and highlights trends such as when they are generally most productive, how often their longest stretch of focus lasts, and more.

How we built it

Intention.ai uses React/Next.js for a responsive web frontend, TypeScript for backend development, and MongoDB for data storage. The system integrates the face-api.js API to enable real-time sentiment recognition via webcam. These signals are paired with the Google Gemini API, which generates personalized, context-aware productivity recommendations. Users can track their emotional state while working, receive tailored strategies to stay focused, and easily control the experience through a Chrome extension toggle that we built. The architecture features JSON Web Token (JWT) authentication, RESTful APIs, and real-time analytics dashboards.

Challenges we ran into

One of the biggest challenges we faced was fine-tuning our AI model to distinguish between three highly similar emotional states—focused, stressed, and tired. Since these states overlap significantly, the model had to capture subtle, nuanced patterns in user behavior. To achieve this, we analyzed micro-signals such as slouching, blinking frequency, and head tilts, mapping these body-language cues to emotional states. Additionally, integrating the live camera feed data into the dashboard was difficult. We had to balance the volume of real-time data collected with the computational cost of processing it, ensuring the system provided high-quality insights without overwhelming resources.

Accomplishments that we're proud of

Achieving real-time emotion detection was a milestone; balancing accuracy with browser performance required creative use of facial landmarks and heuristics. We also take pride in how the Gemini integration turned out. Our early prompts were inconsistent and confusing, but refining context and fallback logic made suggestions feel more useful and user-oriented. The analytics dashboard was another highlight. It pushed us to solve tough challenges in storing and visualizing time-based data, and seeing it come together was especially rewarding. Finally, bringing the backend and authentication together with smooth data flow across extension, web, and database showed we could handle both technical complexity and a polished user experience.

What we learned

Through Intention.ai, we learned how to approach building AI models, skills like combining facial landmark detection with heuristics to balance accuracy and performance in the browser. We grew as full-stack developers, taking a project all the way from a Chrome extension frontend through a Next.js web app to a secure backend with authentication and data analytics. Working with the Google Gemini API taught us how much iteration goes into prompt design, error handling, and fallback logic to make outputs reliable and context-aware (as well as general API implementation). Beyond the technical stack, we also learned how to integrate these pieces into a cohesive system that feels polished and user-friendly.

What's next after PennApps

After the hackathon, we want to refine Intention.ai's functionality further. We want to add a push notification system to provide intermittent, recurring reminders suggesting how the user can best further their study session. We'll also refine the emotion detection model for greater accuracy, expanding the Gemini API integration for more personalized recommendations, and strengthening the dashboard analytics so users can see longer-term trends. We’d like to experiment with different AI model architectures, such as ResNets or a Vision Transformer, to detect more nuances between the different productivity states. Finally, it would be interesting to explore the benefits of gamifying the web application to provide a fun way for users to become more invested in improving their work habits.