A coach that can see the distraction
Most productivity tools wait for the user to report that focus has slipped. Focus Roast reverses that interaction: with permission, it watches periodic screen and webcam snapshots, interprets what is happening, and decides whether an intervention would help.
The tone is intentionally playful. Users choose a Gordon Ramsay or David Goggins-inspired coach, turning a familiar procrastination moment into immediate, character-driven feedback.
Observe
The React interface periodically captures the shared screen and webcam feed.
Describe
A vision model turns those images into a structured account of the current activity.
Decide
A reasoning model judges whether the activity is productive or a distraction.
Respond
The system plays a generated or cached motivational voice response in real time.
Designing around latency
The frontend sends captures to a Flask service, which uses Llama 90B to describe the visual context. DeepSeek R1, accessed through Groq, evaluates that description for procrastination. When feedback is warranted, PlayHT produces the spoken response.
Real-time personality is only convincing when the response arrives quickly. To reduce latency, the system chooses between generating new speech and playing a cached response—saving the expensive path for moments that need specific feedback.
How much context does a coach need to be useful without turning every moment into a slow, expensive model call?