the four instruments
sleuth is a small pipeline of four independent tools, each doing one honest job. No single one is magic; together they're useful.
EXIF / GPS parser
A JPEG stores metadata as TIFF tags. We walk the IFD0 and GPS IFD to pull latitude, longitude, altitude, camera, timestamp and orientation — entirely in the browser, so the raw file never leaves your device.
vision model
A multimodal model on our edge receives a downscaled thumbnail and a strict prompt: return country, region, city, a confidence integer and the visual clues it used. We parse the JSON defensively — if it rambles, we keep the ramble as low-confidence reasoning.
reverse geocoder
Coordinates alone are cold. Nominatim (OpenStreetMap) turns a lat/lon into a display name — a street, a village, a landmark — with no API key and no tracking.
map
A 3D globe (MapLibre GL) over Esri satellite imagery drops the pin on real ground. GPS results get a tight circle; vision estimates get a wide, fuzzy one — the radius itself is the honesty. Scroll to zoom from street level out to the whole planet.
a number you can trust
We compress everything into one integer, 0–100, and we're explicit about what it means.
| score | source | what it means |
|---|---|---|
| 90–100 | GPS in EXIF | coordinates read straight from the file; essentially exact. |
| 50–89 | vision, strong | the model is confident about country and roughly the region. |
| 0–49 | vision, weak | a guess — treat as a lead, not an answer. |
Anything below about 40, we'd rather show you the clues and let you judge, than draw a pin we don't believe in.
"We built sleuth because most geolocation tools confuse confidence with correctness. A coordinate you can't defend is just a lie with a map. So we made the radius the honest part: when we're sure, it's tight; when we're not, it's wide, and we say which one it is."
— Mara Chen, founder · mara c.