Koshika

Offline-first Android app that parses blood report PDFs: extracts biomarkers, tracks health trends, and answers questions via on-device AI.

Description

Koshika (कोशिका - Sanskrit for "cell") is an offline-first Android app that puts health intelligence directly in your hands, no cloud, no API keys, no data leaving your device.

The problem: 200+ million Indians get blood tests every year. They receive

  • PDF reports full of cryptic abbreviations (SGPT, HbA1c, TSH) that they can't

  • interpret. Current apps either require cloud uploads of sensitive health data

  • or don't understand Indian lab formats at all.

What Koshika does:

  • Parses lab report PDFs from Thyrocare, Dr. Lal PathLabs, SRL, Metropolis and more.

  • using regex + a 92-biomarker dictionary with Indian lab name aliases

  • Extracts values, flags out-of-range results, and maps to LOINC codes

  • Visualizes biomarker trends over time with reference range bands

  • Runs an on-device LLM (Gemma via LiteRT) for AI-powered health Q&A

  • grounded in YOUR actual lab data via on-device RAG

  • Exports FHIR R4 bundles compliant with India's ABDM health data standards

Technical depth:

  • ~25MB base APK, PDF parsing + charts work with zero downloads

  • AI features as optional module (~400MB-3GB model download)

  • EmbeddingGemma 308M for on-device semantic search

  • ObjectBox HNSW for vector similarity on mobile

Everything works in airplane mode

Built solo with Flutter. Every inference runs locally. Your health data lives in your cell, not the cloud.

Issues & PRs Board
No issues or pull requests added.