aoisfhdugbos/sanchar-extractor

SancharExtractor — On-Device Immigration Document Field Extraction

A 3.3M parameter encoder-only transformer that extracts structured fields from OCR'd immigration documents. Runs entirely on-device via CoreML (iOS 17+) with zero cloud dependency.

Train on GPU (Free)

Or upload Sanchar_Train.ipynb to Google Colab manually:

1. Go to colab.research.google.com
2. File → Upload notebook → select Sanchar_Train.ipynb
3. Runtime → Change runtime type → T4 GPU
4. Runtime → Run all
5. ~25 minutes to train, downloads model files when done

Architecture

• Model: 4-layer encoder-only transformer, 256 dim, 8 heads
• Parameters: 3.3M
• Vocab: 4,096 tokens (immigration domain)
• Labels: 160 BIO tags covering 79 canonical fields
• Documents: 24 immigration document types (I-797, I-94, EAD, visa stamps, etc.)
• Quantized size: 3.2 MB (INT8)

Quick Validation Results (5K samples)

Metric	Score
Token Accuracy	92.9%
Entity F1	86.5%

Production training on 50K samples with GPU expected to reach 93-96% F1.

Files

├── Sanchar_Train.ipynb          # Complete training notebook (run on Colab)
├── ARCHITECTURE.md              # Full system architecture document
├── data/
│   ├── bio_tags.json            # 160 BIO tag vocabulary
│   ├── form_schemas.json        # 24 document schemas, 355 fields
│   ├── forms_catalog.json       # 97 USCIS forms catalog
│   ├── policy_reference.json    # USCIS policy manual reference
│   ├── edge_cases.json          # 59 immigration edge cases
│   └── reminder_rules.json      # 35 deadline rules + 10 compound rules
├── checkpoints/
│   ├── best/                    # Best validation F1 checkpoint
│   └── final/                   # Final epoch checkpoint
├── tokenizer/
│   ├── sanchar_tokenizer.model  # SentencePiece BPE (4096 vocab)
│   └── sanchar_vocab.json       # Simple tokenizer fallback
└── coreml/
    ├── SancharExtractor_int8.mlpackage  # 3.2 MB, ready for iOS
    └── SancharExtractor_fp32.mlpackage  # 6.2 MB baseline

On-Device Pipeline

Camera → Apple Vision OCR → Document Classifier → SancharExtractor (CoreML) → Structured JSON → Encrypted Vault → Reminder Engine

All PII stays on device. Zero cloud processing.