App interface
TikTok filter

Geo-Safely

Inspiration

We were inspired by how games like GeoGuessr demonstrate that a single photo often contains enough clues to determine its location. A street sign, a building number, or even a flag can reveal more than most people realize. With the rise of AI-powered vision models, these risks only grow — what once took a skilled player minutes can now be done instantly by algorithms. We wanted to flip the script and use AI not to expose locations, but to protect them.

What it does

Geo-Safely is an app that scans your photos for location-revealing objects such as street signs, landmarks, flags, or license plates. Once detected, these areas are blurred, pixelated, or masked so the image can be shared without leaking sensitive context. Users see a side-by-side comparison of the original and redacted photo, giving them confidence about what was hidden.

How we built it

Frontend: React Native app for uploading photos, previewing results, and sharing protected images
Backend: Python FastAPI server with GPU acceleration on an NVIDIA 4060Ti
Models: Gemini 2.5 Flash for suggesting redactions, Grounding DINO for open vocabulary object detection, GeoCLIP for showing how locations can be inferred
Image Processing: OpenCV for applying blur, pixelation, and noise injection filters

Challenges we ran into

Balancing strong privacy filters with keeping images aesthetically usable Very hard to find good models without burning GPU and cloud credits, Grounding DINO was a good find! We spent way too long trying to figure out Lynx and gave up

Accomplishments that we are proud of

Built a fully working end-to-end app in just a few days
Redacted sensitive details in real-world test images while keeping photos shareable
Showed how AI can be repurposed for safety rather than surveillance
Created a tool that also educates users about the hidden risks in their images

What we learned

Privacy is not just about turning off location services. Photos carry context that can be just as revealing
Open vocabulary detection models are powerful but need careful queries for consistency
A privacy tool must be transparent and show exactly what was redacted to earn trust

What's next for Geo-Safely

Extend support to short videos so TikToks and Reels can also be filtered
Optimize inference to run on devices for speed and stronger privacy
Explore applications beyond personal safety, such as journalism, compliance, and law enforcement TIKTOK FILTER