---

AI-Powered Address Transliteration & Validation

An intelligent pipeline designed for e-commerce to bridge the gap between messy user inputs (voice-to-text, regional dialects) and clean, queryable geographic data.

The Problem

E-commerce logistics often fail due to "dirty" address data:

• Phonetic Errors: "Devas" vs "Dewas" (Common in Voice-to-Text).
• Language Barriers: Regional scripts (Hindi/Marathi/etc.) are difficult to query in standard SQL databases.
• Spelling Inconsistency: Non-standard transliteration of local landmarks.

The Solution

This system implements a two-layer validation strategy to ensure high-fidelity data:

1. AI Transliteration Layer: Powered by AI4Bharat Indic Transliteration to map regional scripts to phonetic Roman (English) script.
2. OSM Validation Layer: Cross-references the output with OpenStreetMap (OSM) binary data (.pbf) using Fuzzy String Matching to find the real-world geographic "Ground Truth."

---

Tech Stack

• Backend: FastAPI (Python 3.10)
• Database: PostgreSQL + SQLAlchemy ORM
• NLP: AI4Bharat Indic-Transliteration
• Geo-Data: Osmium & OpenStreetMap (PBF parsing)
• Fuzzy Logic: TheFuzz (Levenshtein Distance)

---

Project Structure

AI-Transliteration/
├── Backend/
│   ├── db.py            # Database connection & Session local
│   ├── models.py        # SQLAlchemy Tables (submissions & normalized_data)
│   ├── schemas.py       # Pydantic models for API validation
│   ├── crud_helper.py   # Core Logic: Transliteration + OSM + Fuzzy Matching
│   ├── translit.py      # AI4Bharat wrapper logic
|   |── main.py
├── osm-data/            # Storage for local .pbf geographic files
├── main.py              # FastAPI entry point & Routes
|── create_db.py
|── ingestion.py
└── requirements.txt     # Project dependencies

---

Installation & Setup

1. Environment Setup

Note: Python 3.10 is strictly recommended due to AI library dependencies.

python3.10 -m venv venv
source venv/bin/activate

pip install "pip<24.1"
pip --version

pip install ai4bharat-transliteration
pip install -r requirements.txt

2. Database Configuration

Update your PostgreSQL credentials in Backend/db.py, then initialize the tables:

python create_db.py

3. Prepare Geographic Data

To maintain privacy and speed, we use local OSM extracts. Place your .pbf file in osm-data/.

# Example: Filtering for Ujjain to optimize performance
osmium tags-filter india-latest.osm.pbf addr:city=Ujjain -o osm-data/ujjain_filtered.osm.pbf

Prepare the orm_lookup table

python ingestion.py

---

Usage

Running the Server

fastapi run main.py

• API Endpoint: http://localhost:8000/submit
• Interactive Docs: http://localhost:8000/docs

Example Request

{
  "pin_code": 456001,
  "state": "मध्य प्रदेश",
  "city": "उज्जैन",
  "locality": "जवाहर मार्ग",
  "landmark": "शांतिलाल मुन्नालाल एंड कंपनी"
}

---

Validation Logic Workflow

1. Ingestion: Receives Hindi/Regional text via FastAPI.
2. Transliteration: AI4Bharat converts जवाहर मार्ग → Jawahar Marg.
3. OSM Retrieval: The system scans local .pbf data for nodes/ways matching the Pincode.
4. Fuzzy Correction: TheFuzz compares Jawahar Marg against all street names in the OSM extract.
5. Normalization: The most likely match (Ground Truth) is saved to the normalized_data table.