nadlens · powered by GeoViT
Visual geolocation with evidence. Istanbul today, Türkiye next, the world after.
Demo coming soon
How it works
nadlens does not rely on one big model trained for one city. It combines signals that work anywhere with local reference data where we have it.
Works everywhere
Street signs, shop names and municipality markers are read and matched against OpenStreetMap to turn text into real places.
Works everywhere
A vision language model looks at architecture, vegetation, road markings and vehicles, and proposes a country, city and area with its reasons.
Where we have reference data
The photo is compared with geotagged street-level images from Mapillary. The closest matches narrow the answer down to a neighbourhood.
Every answer
A final step weighs all signals and returns a location, a confidence circle and the list of evidence that supports it.
A region pack is reference imagery, its embeddings, administrative boundaries and local language hints. Adding a city means adding a pack. The engine stays the same, so coverage can grow from Istanbul to Türkiye to the rest of the world without new model training.
Evidence trail
An illustrative run of the pipeline we are building. The public demo will show this live for your own photos.
> [text] sign detected: "BAHARİYE CD."> [osm] "Bahariye Caddesi" matched in Kadıköy, İstanbul> [clues] tram rails, late 19th century facades, pedestrian street> [region pack: istanbul] 5 nearest references within 400 m> [check] text, clues and references agree> location: Kadıköy, Osmanağa neighbourhood (confidence circle ~300 m)> evidence: 1 street sign, 3 visual clues, 5 reference matchesIllustrative example of the planned pipeline, not a live result.
Region pack #1
The first region pack is built from geotagged street-level images of Istanbul contributed to Mapillary. Each point is a reference location the engine can compare your photo against.
2,298
Reference points shown
39
Districts
Mapillary
Source
Türkiye
Next packs
> loading coverage map...
Use cases
"Where was this photo taken?"
Upload a street photo and get the most likely location, a confidence circle and the evidence behind it.
"Was this photo taken where they say it was?"
Send a photo together with a claimed address. nadlens compares it with that place and tells you whether it fits, and why.
A developer API is planned after the public demo. Join the waitlist and choose "API" to hear first.
Roadmap
No dates yet, just the order. We move to the next phase only when the current one proves itself.
Prove the training-free pipeline beats the v1 model on a fixed Istanbul test set, and measure cost per photo.
Anyone can try it. Istanbul answers at neighbourhood level, other places get a clearly labelled coarse estimate.
Self-serve API with free credits, plus the verify mode for teams that check photos at scale.
Region packs for the largest cities: Ankara, İzmir, Bursa, Antalya and more.
New region packs country by country, measured on public benchmarks.
Responsible use
nadlens is for places. Photos whose main subject is a person are declined.
Answers come as an area with a confidence circle, not an exact point.
Uploads are processed for the request and then discarded.
Reference imagery comes from Mapillary contributors under CC-BY-SA 4.0 and is credited.
Waitlist
We will email you when the public demo opens. No newsletter, no sharing your address.
Where it started
nadlens grew out of GeoViT, an R&D system that trained its own Vision Transformer ensemble on Istanbul street imagery and reached 81-83% accuracy with a four-layer cascade. It worked, but every new city would have meant new training and GPUs. That lesson shaped the training-free design above.
Read the v1 deep divenadcode is a two-person team building its own products.