Doorstep
An open-source local AI walking suggestion application designed for people with limited mobility. It uses the Gemma 3 4B model running locally via Ollama to recommend nearby destinations based on a user's maximum walkable distance for the day, with all AI inference performed on-device to protect privacy.
Timeline (last 90 days)
Doorstep is an open-source local AI walking app designed for people with limited mobility, using a locally running Gemma 3 4B model to recommend walking destinations.
Doorstep's tech stack consists of only Node and a single HTML page; the daemon binds to the local loopback address and generates no outbound network requests beyond the initial download of OSM map data.
Doorstep's core design principle is that the model decides which location to select but never decides how far—distance is filtered at the code level based on accessible range, and the model only chooses from a list that has already been deemed safe.
If the location suggested by the model is not on the filtered list, it is treated as a hallucination rather than a recommendation, and the system falls back to the nearest seating area.
For users with limited mobility, a fabricated destination is worse than no answer, because they will actually walk there.
A developer scraped data from two Indian cities approximately 300 km apart and found: the smaller city had 28 locations, 0 benches, and 3 tagged trees within 1200 m; the larger city had 701 locations, 24 benches, and 566 tagged trees.
A mirror that returns HTTP 200 with no data produces an empty array; because an empty array in JS is truthy it is treated as a success, causing Doorstep to silently save an empty map—this bug originated from a default value of [].
Doorstep deliberately does not implement streaks; instead, it tracks the number of days per month a user goes out and the average accessible distance on those days, with no penalty signal generated for absent days.
The accessible distance combined with precise starting point data held by Doorstep is essentially medical records with home addresses attached; local inference is the foundation of its privacy architecture—with no server, there is no data to be leaked or compelled to be handed over.
All Facts (9)
Doorstep is an open-source local AI walking app designed for people with limited mobility, using a locally running Gemma 3 4B model to recommend walking destinations.
50%UnverifiedDoorstep's tech stack consists of only Node and a single HTML page; the daemon binds to the local loopback address and generates no outbound network requests beyond the initial download of OSM map data.
50%UnverifiedDoorstep's core design principle is that the model decides which location to select but never decides how far—distance is filtered at the code level based on accessible range, and the model only chooses from a list that has already been deemed safe.
50%UnverifiedIf the location suggested by the model is not on the filtered list, it is treated as a hallucination rather than a recommendation, and the system falls back to the nearest seating area.
50%UnverifiedFor users with limited mobility, a fabricated destination is worse than no answer, because they will actually walk there.
50%UnverifiedA developer scraped data from two Indian cities approximately 300 km apart and found: the smaller city had 28 locations, 0 benches, and 3 tagged trees within 1200 m; the larger city had 701 locations, 24 benches, and 566 tagged trees.
50%UnverifiedA mirror that returns HTTP 200 with no data produces an empty array; because an empty array in JS is truthy it is treated as a success, causing Doorstep to silently save an empty map—this bug originated from a default value of [].
50%UnverifiedDoorstep deliberately does not implement streaks; instead, it tracks the number of days per month a user goes out and the average accessible distance on those days, with no penalty signal generated for absent days.
50%UnverifiedThe accessible distance combined with precise starting point data held by Doorstep is essentially medical records with home addresses attached; local inference is the foundation of its privacy architecture—with no server, there is no data to be leaked or compelled to be handed over.
50%