savitr

Fast Surya OCR on Apple Silicon, applied to Indian electoral rolls. savitr runs Surya ~3.6× faster via MLX, and ships a distilled, electoral-roll-specific model — terse-Surya — that emits one compact line per voter (~5× fewer decode tokens at Surya’s accuracy), plus a pipeline that turns scanned roll PDFs into the canonical voter CSV.

Install

pip install savitr               # MLX runtime + terse roll model (auto-downloaded from HF)
pip install "savitr[backend]"    # + the generic MLX Backend for Surya's own pipeline
pip install "savitr[train]"      # + the distillation toolchain

Quickstart

savitr ocr roll.pdf --terse                       # voter records from a PDF
savitr parse-rolls -d english/ -o voters.csv --terse

from savitr import MLXSuryaOCR, parse_terse

eng = MLXSuryaOCR("models/surya-terse-8bit", prompt=...)
voters = parse_terse(eng.ocr_image("page.png")[0])