Bosk (previously Wildr) is a place-based biodiversity app, now in public beta and free to join. Take a photo or record a sound and it suggests what it is with a confidence score; confirm the match and the sighting is saved with its date and place to your personal Wild List. You can also explore a species field guide, a map of nearby sightings and recorded walks. The long-term aim is a map of the living world at the scale of an ordinary city block.
Case Study
Bosk
A field notebook for the natural world: identify animals, plants and fungi from a photo or a sound, log what you find, and see what lives near you.
- Role
- Founder, product & full-stack development
- Year
- 2026
- Status
- Public beta
Now in beta
Species identification, biodiversity datasets and habitat maps all exist, but they rarely come together in one experience built around place. For most people, answering 'what lives around me?' means juggling several apps and databases.
A mobile-first app that combines photo and audio identification with a sightings log and a map of nearby green spaces and water bodies. Identification candidates are always confirmed by the user, and species taxonomy and distribution come from GBIF.
FastAPI · PostgreSQL/PostGIS · pgvector · BioCLIP 2 · BirdNET · Mapbox GL JS · React · Capacitor
Architecture
System Flow
A React and Capacitor client talks to a FastAPI backend over PostgreSQL with PostGIS for geography. Photo ID uses BioCLIP 2 as an embedding model: an image embedding is compared against a regional species-embedding table in pgvector and the top candidates are returned. Bird audio goes through BirdNET. GBIF backs taxonomy and distribution, rolled out region by region, starting with Berlin and Germany.
Learnings
What This Work Sharpened
- A similarity score is not accuracy: in testing, a 0.85 cosine similarity was only about 77% correct, so thresholds have to be calibrated on real evaluation data
- When a data provider closed API access, designing the identification layer around swappable engines kept the product moving
- Staging the reference data by region keeps the model general while setting honest expectations about local accuracy