Teffen EllisFOSS4G:UK 2026 · Leedshttps://mailwoman.ai

Truth orConsequencesNew MexicoApproximately Correct

A brief talk on address parsers, machine learning, and geocoders.

A 30-minute talk at FOSS4G:UK 2026 in Leeds about Mailwoman: a postal-address parser built as a compact neural sequence labeler instead of hand-written rules or a large language model.

Session

When

Tuesday 13 October 2026, 15:30–16:00

Where

Room "Think", FOSS4G:UK 2026, Leeds

Programme

Talk page on talks.osgeo.org

What it covers

• Why a small sequence-labelling model, rather than a rule engine or an LLM, for postal addresses.
• Reference data — postal registries, gazetteers, postcode indexes — feeding the model as soft evidence instead of overriding it.
• One shared model of about 40 MB, int8-quantized ONNX, with SentencePiece tokenization and BIO decoding constrained by transition masks and Viterbi.
• Per-locale overlays of a few hundred kilobytes, with a walkthrough of building one for en-GB.
• The same weights running in Node.js and in the browser over WebGPU or WASM, next to a SQLite-WASM gazetteer, so parsing and geocoding can stay on the client.
• Drop-in compatibility with Nominatim-, Photon- and libpostal-style APIs.
• Training and resolution data from OpenStreetMap, Overture, Who's on First, and national registers such as BAN, LINZ and TIGER.

An earlier Mailwoman talk: Paris, Texas: Technically Correct.