Errand Walker

Errand Walker turns a messy to-do list into one walk. Type the list the way you would type it into a notes app, for example pharmacy, post office, buy rice, and the app answers in a single line: how many stops, how long on foot, how long by car, and whether you can skip the car. It is a phone-first web app for anyone with a handful of small errands in a place with decent OpenStreetMap coverage, and it also understands some Indonesian input.

A small open-weight model does exactly one job: Gemma 3 1B, run by llama.cpp, maps each errand to one of ten fixed categories (pharmacy, post office, groceries, ATM, bakery, hardware, laundry, cafe, clinic, market) or other, with output constrained to a JSON schema. OpenStreetMap data via Overpass finds the nearest matching places and OpenRouteService routes one loop. Three deterministic layers sit around the model: a keyword parser with fuzzy matching that overrides the model on a clear match, a grounding check that drops any model errand sharing no word with what was typed, and a fallback to the keyword parser alone if the model is down. The model needs no API key and no per-call cost, and with the bundled model no AI vendor sees the errand text.

The author measured the parser honestly: on a dev set of 26 cases (57 expected errands) the raw model scored 82.5 percent, and on a held-out set of 15 cases (26 errands) 96.2 percent, with the caveat that the held-out cases were shorter and simpler. Vibe-coded evidence, in the author's own words on DEV (11 October 2026): "I built this with Claude Code." Note: the post's README excerpt also says the app was built with an AI coding agent (OpenCode, model claude-sonnet-5-5) from a human-written product spec, so the build involved both framings; aiTools records Claude Code per the explicit first-person sentence in the post's AI use section. GitHub stars: 0 at the 28 October 2026 check.

Honest limits: the author did not complete a timed walk to compare predicted and real minutes, so the minute figures are predictions, not measured walks. Location does leave the device: the server sends the start point rounded to three decimals to Overpass and the exact start to OpenRouteService, though the server does not store or log coordinates or errand text and the app has no analytics and no cookies. OpenStreetMap quality is the real ceiling: in one smoke test, buy rice resolved to a batik shop tagged as a general goods store. The demo runs on Render (1 CPU / 2 GB) and the first load after a quiet spell is slow while the model loads; demo mode uses a fixed public landmark in central Yogyakarta, Indonesia, so demo requests carry that point, not yours.