AI audio analysis platform for automated cough detection and classification

The project: a production-grade platform for AI audio analysis

Resmonics engaged us to build a scalable, secure and maintainable platform for automated cough analysis. The challenge was taking an existing machine learning pipeline into a robust production environment that accepts audio uploads, orchestrates several inference engines, and stores the results in a structured, data-protection-compliant way.

The solution: microservices architecture and ML integration

devs group designed and built the entire backend as a microservices architecture. Our technical expertise produced a single coherent system that orchestrates several model formats under one roof and secures the quality, scalability and maintainability of the whole.

API gateway: the central entry point for audio uploads, authentication (OAuth2 / OpenID Connect) and recording management.

ML inference: three inference engines run inside one unified pipeline.

Persistence and API: robust persistence in PostgreSQL with versioned schema migrations, clearly defined interfaces based on OpenAPI, and compliant storage (hashed IP geolocation).

DevOps and infrastructure: operated and delivered through a modern container and CI/CD landscape with Docker and a reverse proxy, versioned through Git tags.

Piloting

For a pilot phase we additionally built a dedicated frontend, so the solution could be tested and validated in real use. After a successful pilot, the platform went into production as a pure API solution, with an international pharmaceutical group as the end client.

The result: a scalable, compliant analysis platform

Building the platform deliberately and taking the pipeline into production produced a capable, stable system. The solution is marked by a clean hexagonal architecture, high technical quality, and data protection by design.

Frequently asked about this project

What was the hardest part?

The route from research into production. Resmonics’ machine learning pipeline worked, but a model in a lab is not yet a platform that accepts audio uploads securely, orchestrates three inference engines, and stores results in a compliant way.

How is data protection handled?

Through architecture rather than policy: authentication over OAuth2/OpenID Connect, clearly defined OpenAPI interfaces, and storage that keeps things like IP geolocation only in hashed form. Health-adjacent audio data demands that care, and demands it from the first architecture sketch rather than as a retrofit.

Why several inference engines?

Because several model formats are in use and they keep evolving. The unified pipeline runs all three engines under one roof, which makes a new model a deployment rather than a rebuild of the platform. For Resmonics that means research can keep developing models without the platform being touched each time, and the platform stays stable while the models beneath it move.