Healthcare • LLM • Audio AI

[REDACTED] US HealthTech

HIPAA-Compliant Clinical Documentation Assistant

Client:US-Based HealthTech SaaS (New York)
Timeline:8 months (2024)
Team:AI Engineering, Backend & Web
Mistral 7BWhisperFHIR APIHIPAAEHR IntegrationTensorRT
System Architecture

The Challenge

A US healthtech SaaS serving 800+ independent physician practices faced client attrition as competitors launched AI-assisted features. Their physicians were spending 40% of their day on manual note-writing (SOAP notes, discharge summaries), significantly impacting patient care and revenue.

  • Physician burnout from excessive administrative burden
  • High PHI exposure risk using external, third-party AI APIs
  • Inconsistent note quality and medical coding accuracy
  • Need for seamless integration with legacy EHR systems (Epic, Athenahealth)

The Solution

Developed a fully on-premise clinical documentation assistant that transcribes and structures physician-patient interactions into medical notes within the client's secure VPC.

  • On-Premise Clinical LLM (Mistral 7B fine-tuned on clinical corpora)
  • Real-Time Medical Transcription Pipeline using a fine-tuned Whisper model
  • EHR Integration Layer with bidirectional HL7 FHIR API connectors
  • Physician Documentation Interface (React) for inline editing and submission
  • AI Quality Monitoring Dashboard tracking physician acceptance and model drift

Lexer System's Approach

1

Mistral 7B Fine-Tuning

Optimized Mistral 7B on de-identified clinical notes and ICD-10/CPT coding guidelines to ensure medical accuracy and structured output consistency.

2

Whisper Pipeline Optimization

Fine-tuned OpenAI's Whisper for medical terminology and ambient clinical environments, ensuring high-accuracy transcription in noisy exam rooms.

3

HL7 FHIR Integration

Built secure connectors to Epic and Athenahealth to pull patient context for note generation and push finalized notes back into the EHR.

4

NVIDIA TensorRT Deployment

Utilized TensorRT for model quantization and high-speed inference, allowing the system to run efficiently on the client's private cloud GPUs.

Results & Impact

18m → 4m
Documentation Time

Average time spent per patient note

78%
AI Acceptance Rate

Notes accepted by physicians with zero or minor edits

96.2%
Coding Accuracy

Precision in automated ICD-10 and CPT code suggestions

HIPAA Certified
Compliance

Passed third-party HIPAA audit with zero PHI leakage

Technical Highlights

Self-Hosted Privacy

A completely isolated AI stack where no voice data or clinical text ever leaves the physician's secure environment.

Medical Audio Processing

Specialized denoising and speaker diarization to separate physician instructions from patient responses.

Context-Aware Notes

The LLM uses existing EHR patient history to provide context-aware summaries that go beyond simple transcription.

Lessons Learned

  • On-premise deployment is the key to healthcare AI adoption; physicians do not trust external APIs with PHI
  • Model fine-tuning on medical coding standards is essential for accurate billing integration
  • Ambient audio recording is more valuable to physicians than dictated notes

Next Steps

  • Expand to multi-lingual transcription for diverse patient populations
  • Implement automated referral letter generation based on SOAP notes
  • Integrate real-time clinical decision support during the documentation flow

Have a Similar Challenge?

We specialize in building production-grade systems that solve complex operational problems. Let's discuss how we can help architect your solution.