Anchor

Camera Settings
Gemini detecting me being distracted
Timeline of distractions and configuration options
Ending session early

Inspiration

As university students, we’ve seen how traditional productivity tools often fail neurodivergent users. Many existing proctoring or "focus" apps feel invasive and punitive, treating a loss of focus as a moral failing rather than a hurdle in executive function. We wanted to build Anchor—not as a digital cage, but as a cognitive prosthetic that provides the external inhibition and structure that ADHD and other neurodivergent brains often struggle to maintain internally.

What We Built

Anchor is a browser-integrated focus ecosystem designed for deep work. It moves beyond simple site blocking by using multimodal AI to understand the context of your work.

Context-Aware Monitoring: Uses Gemini (Google AI Studio) to distinguish between "productive" and "distracting" content, even on the same platform (e.g., a YouTube tutorial vs. a YouTube rabbit hole).

Physical Presence Tracking: Employs head-pose estimation to detect when a user has physically drifted away from their task.

Proactive Intervention: Redirects users to their stated intentions the moment a distraction is detected.

How It Works

The system operates on a dual-engine architecture:

Local Computer Vision: A webcam feed processes head orientation (yaw and pitch) locally to ensure privacy. If attention drops below a threshold, the UI provides immediate visual feedback.
Cloud-Based Semantic Analysis: Using the Gemini API, the extension captures periodic screen fragments. Gemini analyzes the visual data against the user’s "Stated Goal." If the AI detects a context mismatch (e.g., the user is browsing clothes while their goal is "Write Engineering Report"), it triggers a redirection.

Challenges & Learnings

The primary challenge was minimizing latency while using a high-level LLM for vision tasks. By optimizing the interval of Gemini calls and handling site-blocking via Manifest V3 declarativeNetRequest APIs, we achieved a "snappy" feel that is crucial for maintaining a user's flow state. We learned that for neurodivergent users, the speed of the feedback loop is just as important as the accuracy of the detection.