All work
122024SCAD

Smart Form Validator (Arabic Address & Name Parsing)

Standardising messy Arabic names and addresses in-line — lifting record-linkage match rate from 72% to 94%.

Before

Field-collected forms had inconsistent Arabic name and address formatting that broke downstream record linkage.

After

A light GPT-3.5 normaliser with a transliteration model standardises names, splits address components, and validates against the national address registry — in-line on submit at sub-200ms p95. Match rate rose from 72% to 94% and eight brittle regex rule-sets were replaced by one model plus a small ruleset.

Challenge

Field-collected forms had inconsistent Arabic name and address formatting that broke downstream record linkage.

Approach

Light GPT-3.5 normaliser + transliteration model that standardises names, splits address components, and validates them against the national address registry.

How it was built

  1. 1

    Normalisation model

    Phase 1

    Built a light normaliser + transliteration step to standardise Arabic names and split address components consistently.

  2. 2

    Registry validation

    Phase 2

    Validated standardised addresses against the national address registry so downstream linkage has clean, verified inputs.

  3. 3

    In-line at submit

    Phase 3

    Ran the whole thing in-line on form submit at sub-200ms p95 — fixing data at the source instead of cleaning it later.

Key architecture decisions

Fix data at entry, not downstream

Why · Cleaning at submit time stops bad records from ever entering the pipeline, which is why linkage jumped 22 points.

One model + small ruleset over 8 regex systems

Why · The brittle regex rule-sets were unmaintainable. A model plus a thin ruleset was more accurate and far easier to keep working.

Impact

  • Improved record-linkage match rate from 72% to 94%
  • Replaced 8 brittle regex rule sets with one model + small ruleset
  • Runs in-line on form submit, sub-200ms p95 latency
+22pp
match
200ms
latency
-8 systems
rules

What I'd tell someone building this

  • 01 · The cheapest place to fix data quality is the moment of entry.
  • 02 · A small model can retire a pile of brittle regex — and be easier to maintain.
  • 03 · Latency budgets matter for in-line validation; sub-200ms keeps it invisible to users.

Tech stack

GPT-3.5Azure FunctionsArabic Transliteration

Ask anything about Smart Form Validator (Arabic Address & Name Parsing)

AI scoped to this project · GPT-OSS 120B