Case Study

How our Claude Developers Built a Remote Diagnostics System for a Healthcare ISV

TL;DR: Claude on Bedrock for healthcare powers SaluberMD's remote diagnostics pipeline. GoML's Claude developers shipped it in 8 weeks on a HIPAA-compliant AWS stack. Outcomes: 99% accuracy on auscultation audio classification, 95.71% accuracy on standardised medical image classification, and 80% clearer diagnostic audio after enhancement. Result: 90% faster clinical AI deployment for a live telehealth product.

For telehealth companies, remote diagnostics can't be a "nice to have." Clinicians using handheld devices to capture heart sounds, lung sounds, and medical images need AI that runs in real time — without breaking HIPAA or clinician trust.

SaluberMD came to GoML looking for that foundation. Eight weeks later, they had it in production, on Claude on Bedrock for healthcare workloads.

The Problem

Remote diagnostics generate large volumes of unstructured clinical audio and image data. In a telemedicine setting, that data is inherently messy.

  • Audio quality varies by device, environment, and how the clinician holds the microphone.
  • Body-part tags on images are often wrong or missing, forcing manual review.
  • No standardised preprocessing pipeline meant every downstream model was fighting the data before it could interpret it.

On top of that, SaluberMD needed all of it exposed as low-latency APIs, deployed on HIPAA- and GDPR-compliant AWS infrastructure, and integrated with the existing telemedicine backend.

The combined constraint set is what usually kills clinical AI pilots. Getting through it is the whole job.

The Solution: An Agentic Pipeline on Claude on Bedrock

GoML's Claude developers built the diagnostic pipeline as an 8-week engagement. Two layers, one contract.

Layer 1 — Specialised Models for Signal Work

PyTorch models trained for audio enhancement, auscultation classification, image classification by body part, and disease detection.

Layer 2 — Agentic Orchestration on Claude on Bedrock

The agentic AI layer autonomously manages preprocessing, validation, classification, and inference across incoming clinical data. This is what removes the manual step from the middle of the diagnostic loop.

Everything runs behind secure REST APIs on a healthcare-compliant AWS architecture, integrated with SaluberMD's existing telemedicine platform.

What the Pipeline Does

AI-Powered Audio Enhancement

Noise reduction and signal isolation for heart and lung sounds. Clinically clean WAV outputs a physician can actually review. Works across the range of mobile capture conditions telemedicine sees in the field.

Auscultation Audio Classification

Validates recordings against the labelled body part (heart, lungs, bowel). Returns structured JSON with confidence scores, so clinical reliability is inspectable rather than assumed.

Image Classification and Disease Detection

Body-part classification for ear, throat, and skin. Disease classification for skin (nine classes). Standardised preprocessing pipelines that let SaluberMD extend to additional organs and conditions without rebuilding the ingest layer.

Developer-Ready APIs and a Validation UI

Secure REST APIs for backend integration. Streamlit validation UI for clinicians and QA teams doing quality assurance on the agentic pipeline itself.

The Impact

  • 99% accuracy in auscultation audio classification
  • 95.71% accuracy in standardised medical image classification
  • 80% clearer diagnostic audio after enhancement — meaningful improvement in clinical signal quality
  • Near real-time clinician feedback loop replacing delayed manual validation
  • 8-week production timeline from scope to live pipeline

Before vs After

AspectBeforeAfter
Diagnostic data qualityNoisy audio, inconsistent imagesAI-enhanced, standardised clinical inputs
Audio validationManual review of auscultation recordingsAutomated classification with confidence scores
Image processingManual validation and preprocessingStandardised AI-driven preprocessing and classification
Clinician feedbackDelayed, inconsistentNear real-time, AI-assisted validation
ScalabilityLimited by manual workflowsAPI-driven, cloud-native, elastic

"With this pipeline live, we replaced manual validation and inconsistent preprocessing with a reliable, low-latency diagnostic foundation. Clinicians now get cleaner signals, faster feedback, and greater confidence in virtual care decisions."

— Prashanna Rao, Head of Engineering, GoML

The Stack

AWS · Amazon API Gateway · Amazon EC2 · Amazon Bedrock (Claude) · PyTorch · FastAPI · Streamlit · IAM · TLS · Amazon S3

Key Takeaways for Telemedicine Platforms

Challenges You'll Hit

  • Unstructured audio and image data undermining diagnostic reliability
  • Manual preprocessing bottlenecking clinical workflows
  • Lack of standardisation capping how far a telemedicine AI stack can scale

How to Move Past Them

  • Use agentic AI — Claude on Bedrock for healthcare workloads works well — to automate preprocessing, validation, and inference
  • Standardise clinical audio and image inputs before model execution
  • Expose everything through secure, low-latency APIs
  • Build on healthcare-grade cloud infrastructure so compliance isn't a retrofit

Build the Same Foundation for Your Telehealth Product

If your telemedicine platform is stuck at manual validation and inconsistent data quality, an 8-week engagement with GoML's Claude developers can put you on the same footing SaluberMD is on today.

Speak with Our Claude Developers

The planet's best Claude developers can be working with you next week. Tell us your needs and we'll get in touch right away.