Skip to content

Echo: Ambient Clinical Intelligence

Core Problem

Healthcare systems are burdened by documentation overload. We are losing physicians to administrative burnout: doctors spend on average two hours managing Electronic Medical Records (EMR) for every one hour spent listening to patients.

For patients, the current system is equally broken. Medical instructions are hard to follow, and the formal clinical language generated by classic hospital systems creates a wall of confusion and anxiety.

While AI "scribes" exist, they lack the one feature absolutely required in medicine: Trust. Standard wrappers over legacy language models hallucinate facts. In healthcare, hallucinating a diagnosis or a drug dosage is dangerous. Furthermore, transcribing and streaming raw Protected Health Information (PHI/PII) to unverified LLM endpoints violates strict privacy frameworks.

Echo Solution

Echo - Medical AI Companion is a deterministic, compliance-first ambient engine. It is designed to completely eliminate documentation labor while returning agency and understanding to the patient.

Echo listens passively to the doctor-patient dialogue and automatically produces twin documents:

  1. Structured EMR Report: A strictly typed JSON representation, output to formal PDF format, ready to be ingested by the hospital database.
  2. Simplified Patient App: A translated, layman-friendly summary sent instantly to the patient's companion application.

Technical Foundation

Echo is built upon two distinct engineering pillars:

1. Zero-Hallucination Pipeline

Echo text extraction utilizes the native OpenAI Structured Outputs API to enforce strict Pydantic parsing natively within the LLM inference step. We do not prompt the model to generate free-text prose. Instead, we force the AI to produce isolated clinical facts. Every single extracted fact must mathematically cite a perfect string match (an "exact quote") from the raw audio transcript. The Python backend then runs a deterministic substring check against the transcript. If the quote doesn't exist, the extraction is rejected. This guarantees total auditability.

2. Opaque Pointer Architecture

Patient data privacy is guaranteed. General healthcare records from the national EESZT database are never fully loaded into our backend. Instead, we utilize Opaque Pointers—generic document identifiers—that Echo identifies during speech. These IDs are rendered on the patient's device as secure portal hyperlinks, ensuring the AI itself never ingests historical personal records.

Quick Start

To run the Echo platform locally:

git clone https://github.com/GregoJaca/GDE-MIT-mesh.git
cd GDE-MIT-mesh
./start.sh

Navigate to http://localhost:5173.