Sister Software

Cheap structural priors for the runtime pipeline: character class, segmentation, known-format detection. Pure functions, no ML.

Mailwoman core: tokenization, decoder, formatter, resources, filters, utils.

Stage 2 of the runtime pipeline — derive a LocaleHint from QueryShape (script + known-format hits). Rule-based v1; trained classifier deferred to v0.6.0.

Stage 2 of the runtime pipeline — derive a LocaleHint from QueryShape (script + known-format hits). Rule-based v1; trained classifier deferred to v0.6.0.

Stage 1 of the runtime pipeline — deterministic input preprocessing (Unicode NFC, punctuation, whitespace, abbreviation). Pure functions, no ML.

Mailwoman corpus pipeline: BIO-labeled dataset builder for the neural classifier.

The geocode-first record matcher: block → score → cluster. This first reduce ships the string comparators (Jaro / Jaro-Winkler + an edit-distance fallback for compound surnames) that the Fellegi-Sunter scorer is built on.

Stage 2.7 of the runtime pipeline — propose coherent input units (boundary discovery) with a structural kind hypothesis + confidence. Rule-based v1 (port of v1 section/sub-section logic); learned 1-2M-param span proposer reserved for v0.5.1.

The geocode-first record-matching application: resolve messy contact/organization records into canonical, geocoded entities (block → score → cluster) and export them as GeoJSON for spatial analysis. The clinic-funding use case mailwoman was built for.

Mapping utilities for the Mailwoman geo-spatial data pipeline.

Mailwoman neural-classifier weights for locale 'en-us'. Data-only package — loaded by @mailwoman/neural at runtime.

Mailwoman neural-classifier weights for locale 'en-nz'. Data-only overlay — shares the base model.onnx + tokenizer.model from @mailwoman/neural-weights-en-us (they are byte-identical); ships only the nz-specific data siblings. Loaded by @mailwoman/neural

Mailwoman neural-classifier weights for locale 'en-gb'. Data-only overlay — shares the base model.onnx + tokenizer.model from @mailwoman/neural-weights-en-us (they are byte-identical); ships only the gb-specific data siblings — pair-index-gb.bin and postc

Environment Agency Flood Map for Planning (England) — acquisition, build, and the flood.db layer reader

A calibrated, retrieval-augmented postal-address parser — CLI + library.

zh-CN weights overlay for mailwoman — data-only; shares the base model with @mailwoman/neural-weights-cjk.

ja-JP weights overlay for mailwoman — data-only; shares the base model with @mailwoman/neural-weights-cjk.

Mailwoman neural classifier runtime: SentencePiece tokenizer + ONNX inference + decoder wiring.

es-ES weights overlay for mailwoman — data-only; shares the base model with @mailwoman/neural-weights-en-us.

it-IT weights overlay for mailwoman — data-only; shares the base model with @mailwoman/neural-weights-en-us.

German (de-DE) weights overlay for mailwoman — data-only; shares the base model with @mailwoman/neural-weights-en-us and ships the DE placetype-pair index.

Mailwoman neural-classifier weights for CJK scripts — the char-path base model (Japanese + Chinese under one head). Data-only package: char-vocab.json + model-card.json at 0.0.1; model.onnx joins the shipped files with the first functional release. No Sen

Indian English (en-IN) weights overlay for mailwoman — data-only; shares the base model with @mailwoman/neural-weights-en-us and ships the IN placetype-pair index.

Mailwoman neural-classifier weights for locale 'en-au'. Data-only overlay — shares the base model.onnx + tokenizer.model from @mailwoman/neural-weights-en-us (they are byte-identical); ships only the au-specific data siblings. Loaded by @mailwoman/neural

DEPRECATED — the browser neural runtime moved into @mailwoman/neural. This package re-exports it so existing imports keep working; migrate to @mailwoman/neural/web-loader and @mailwoman/neural/web-onnx-runner.

Stage 2.5 of the runtime pipeline — categorize inputs by query shape (postcode_only / locality_only / structured_address / intersection / po_box / landmark / vague). Rule-based for v1.

The annotations composer — a native typed AnnotationSet, the Annotator interface, and OpenCage/native serializers. Each recipe package (timezone, un-locode, coordinate formats, country reference) implements Annotator; this composes them into the block the

Coordinate → IANA timezone, server-side. Point-in-polygon over timezone-boundary-builder polygons in a node:sqlite DB; UTC offset via Intl (no tz-db dependency). An @mailwoman/annotations Annotator.