SEHAT AI (System for Evidence-based Healthcare Assessment & Triage)

SEHAT AI turns a patient's voice, lab-report photos, and medical images into a source-linked triage note that a human reviews in seconds. It extracts and summarizes but relies entirely on hardcoded medical protocols for urgency, so it never diagnoses.

By the finale, we will demo one patient's journey end-to-end, on a single laptop, mostly offline: ->Multilingual intake: Consent and voice intake in Hindi, Odia, or English using on-device speech recognition (Silero VAD + IndicConformer) with a read-back confirmation step. ->Document & image understanding: On-device OCR (PaddleOCR + Surya + Chandra) reads lab reports and prescriptions, cross-checking values against a local RxNorm database. X-rays and ECGs get an AI-described visual findings card (using MedGemma 1.5) strictly labeled as "not a diagnosis". ->Source-linked summary: An LLM (Microsoft Semantic Kernel adapter) produces a triage note where every single field carries its exact quote, source location, and a 3-pass agreement score to eliminate hallucination. ->Reviewer dashboard: A deterministic rules engine sets RED/YELLOW/GREEN urgency based on AIIMS/NEWS2/qSOFA protocols. Cases appear in a priority queue with a RED review-time target. ->Referral note generator: A one-click handoff letter auto-filled from the reviewed note, tracking transport and destination, with a FHIR R4 bundle export for ABDM-style exchange. ->DPDP Privacy Pack: Purpose-wise consent, PII redaction (Presidio), NeMo Guardrails, and a hash-chained audit log for supervisor verification.

Hardware: One laptop (16 GB RAM) to simulate Edge AI deployment at rural PHCs. Runs fully offline with optional Azure cloud fallback. Tech Stack: Backend: Python FastAPI, SQLite (aiosqlite), Pydantic. Frontend: Next.js 15, React 19, Tailwind 4. Orchestration: Microsoft Semantic Kernel (aligned for Cognizant TriZetto AI Gateway). AI/LLM: Azure OpenAI (Cloud) & Qwen3-4B-Instruct via llama.cpp (Offline). Safety Core: Pure-Python rules engine (AIIMS Protocol, NEWS2, qSOFA). Voice & Translation: Silero VAD, AI4Bharat IndicConformer, IndicTrans2. OCR & Vision: PaddleOCR, Surya OCR 2, Chandra OCR 2, local RxNorm, MedGemma 1.5. Privacy: Microsoft Presidio, spaCy, NVIDIA NeMo Guardrails.

We have merged 8 build phases with 1,255 backend tests and 88 frontend tests passing on synthetic data: Safety core: Built the deterministic rules engine covering AIIMS Triage Protocol, NEWS2, qSOFA, and 7 scenario packs. A test enforces that this engine contains no LLM or network calls. Voice: On-device Hindi speech recognition runs completely offline. Browser walkthroughs with microphone, read-back, and prefill steps are verified. OCR: Local OCR pipeline runs on synthetic lab reports/prescriptions, highlighting values on the page and utilizing a dispute flag when our 2 OCR engines disagree. LLM summary (live, local): Qwen3-4B processed 60/60 synthetic cases successfully. Out of 15 red flags, the model alone caught 0, but paired with our deterministic keyword layer, the system caught 15/15 with zero false positives. Medical images: Smoke tested MedGemma 1.5 local parsing for 4 image types. Reviewer dashboard & Privacy: Built rules-ordered queue, RED review targets, sign-off codes, purpose-wise consent, PII redaction, and a verifiable hash-chained audit log.