Conscience AI

our main page

Inspiration

We're all horrible at getting things done. Even when we want to focus, it's far too easy to get sidetracked—whether it’s a phone notification, a Wikipedia rabbit hole, or a “quick” League of Legends match that turns into an hour. Traditional productivity tools rely on self-discipline; Conscience AI adds something stronger: accountability.

What it does

Tell Conscience AI your goal, and it will start a focus session, monitoring your screen and attention in real time. The moment you drift onto something unproductive, it nudges you—literally talking you back into focus. It’s like having a tiny, relentless mentor sitting on your shoulder, making sure you actually follow through on what you intended to do.

How we built it

We combined:

On-device screen monitoring to extract high-level context and detect task drift without storing user data.
Lightweight attention tracking using webcam signals to estimate focus loss.
A real-time feedback loop built with an LLM that generates gentle (or firm) voice nudges when the user falls off task.

Challenges we ran into

Balancing accountability with privacy: We didn’t want to send any screenshots or webcam data to external servers, so we engineered everything to run locally.
False positives: Early models kept yelling at us even when we were working.
Smooth voice interaction: Creating nudges that were effective without feeling annoying—or creepy—required lots of iteration.
Calibration for different working styles: Some users multitask, some deep-focus; designing an adaptive system was harder than expected.
Background noise reduction: Early models were too responsive to background noise. It took many iterations for models to respond to exactly what we wanted them to.

Accomplishments that we're proud of

Built a fully functional, on-device, real-time accountability system in under 48 hours.
Created a surprisingly effective attention-tracking pipeline with minimal hardware requirements.
Verified that our model significantly reduced distraction events in a small user study.

What we learned

Accountability matters more than fancy productivity features—humans work better when something is watching (in a non-creepy way).
Attention is messy and extremely personal; building a one-size-fits-all model doesn’t work.
Small, immediate feedback is disproportionately powerful—just a single voice cue can break a distraction loop.
Privacy-by-design is essential. Users trust tools that respect boundaries.

What's next for Conscience AI

Adaptive coaching styles: users can choose “supportive,” “strict,” or “dead serious” modes.
Integrations with calendars, notebooks, and task managers to automatically start focus sessions.
Mobile app companion for phone distraction detection.
Deeper analytics: heatmaps of distractive patterns, personalized focus recommendations.
End-to-end encryption + local LLM options to strengthen privacy and performance.