Inspiration

Compulsive use of adult content is a real struggle for many people, but almost every existing tool is just a website blocklist — it only blocks known URLs, can't catch images, and does nothing in the exact moment the urge hits. We wanted something that doesn't just block, but actually catches the person when they're about to slip.

What it does

LAVOCADO is a local-first desktop guardian. It watches each display in real time, detects explicit content on screen, and instantly covers only the affected screen. In that moment, instead of a cold warning, an AI companion gently guides the user through the urge with a short, personalized typing meditation.

Two things make it different:

Block + Catch: computer vision blocks the content; an AI companion helps the person ride out the urge.
Privacy first: screenshots never leave the device. Detection runs 100% locally, and only a supportive-message request is ever sent to the AI — never screenshots, URLs, window titles, or detection labels.

How we built it

Vision layer: NudeNet 640m at 640px inference, with per-body-part candidate and strong-evidence thresholds, plus 2-of-3 temporal confirmation to cut false positives. Small-content rescue splits each monitor into tiles for a second look.
Intervention layer: when content is confirmed, the affected display moves through Pause → Breathe → Ready. In the Ready stage, an LLM asks a few short questions, then guides an interactive typing meditation tailored to the time of day and how many times the user was interrupted today.
Privacy & storage: only minimal event metadata (time, label, confidence, monitor, whether intervention was shown) is stored in a local SQLite database. No screenshots, URLs, or window titles are ever stored or uploaded.
Cross-platform: Windows, macOS, and Linux, with platform-native APIs isolated behind a single adapter.

Challenges we ran into

Balancing detection: too sensitive floods users with false positives on swimwear or art; too loose misses real content. We solved it with per-part thresholds and multi-frame confirmation.
Privacy vs. AI: keeping everything local while still using an LLM — so we send only a fixed supportive-message request, never any content.
Helping, not shaming: we designed the intervention to be invitational and never punitive — the user can always exit, so it supports rather than punishes.

Accomplishments that we're proud of

A working block-and-catch loop: real-time local detection, instant per-display cover, and an AI companion that actually helps in the moment.
A privacy design where the screen never leaves the device — we can even demo the product without ever showing sensitive content, because the cover hides it.
Deliberate restraint: we only block clear, objective content instead of chasing subjective edge cases that would flood users with false positives.

What we learned

The hardest part wasn't detection — it was restraint. A guardian that over-blocks gets uninstalled on day one. Choosing to block only clear content, and handling the personal, subjective side through behavior and conversation, is what makes it something people can actually keep using.

What's next for Lavocado

Long-term memory so the companion learns each user's patterns and intervenes earlier, plus smarter behavior-based prevention that steps in before the urge even peaks.