We rebuilt a global healthcare provider's telehealth video platform on distributed video servers (SFUs) that scale with demand and keep patient data out of the logs. It supports 50k+ concurrent sessions with 99.99% uptime.
Industry: HealthcareService: SaaS & SoftwareUpdated
Overview
The challenge
The old telehealth platform sent every call through one large peer-to-peer signalling server. It crashed during the morning peak, between 8 AM and 10 AM EST. With 2M+ patients, even a 1% failure rate meant thousands of missed appointments and frustrated doctors.
Patients on weak rural connections had poor calls even when the server stayed up, and the logs could not show who had accessed call data. The provider needed more capacity and a clear privacy record at the same time.
High packet loss on rural and unstable networks made diagnostic video calls impossible to read.
Leading specialists were quitting the platform because calls failed during their busiest slots.
The old logs lacked the traceable privacy records needed under business associate agreements (BAAs).
The solution
We set up Selective Forwarding Units (SFUs) in 12 AWS regions, so each call takes the shortest network path between doctor and patient.
The cluster scales on the number of active consultation rooms, using custom Horizontal Pod Autoscaler metrics, rather than on general CPU load.
Every signalling event and log is scrubbed of patient health information before it is saved to a traceable audit trail.
Before and after
Before
Single-server limit
The platform crashed at 5,000 concurrent users when the signalling server slowed down.
After
Scales out by region
Regional load balancing supports 50,000+ sessions at the same time.
Before
No access record
There was no consistent log of who accessed signalling data, or when.
After
Full audit trail
Traceable logs and end-to-end encryption with rotating room keys.
Before
Calls dropped
Switching from Wi-Fi to 5G often cut off a consultation.
After
Calls recover
ICE restart and packet-loss concealment keep the session running through the switch.
Key features
The platform checks network quality in real time and lowers video resolution when needed, without dropping the audio the doctor relies on.
Business impact
Zero dropped diagnoses
Performance metrics are collected without ever exposing patient health information to the analytics layer, in line with the provider's BAAs.
Business impact
No patient data in analytics
New features go to 1% of traffic first, so a faulty release cannot take down calls for every patient.
Business impact
99.99% availability
Results
10x Capacity Lift
Automatic scaling lets the platform absorb the morning peak without calls slowing down.
Before: 5k UsersAfter: 50k+ Users
60% Response Time Reduction
Routing calls through the nearest region cut delay between video packets, so specialists get a clear, steady picture.
Before: 240msAfter: 95ms
99.9% Success Rate
ICE restart logic stopped calls dropping when a patient's phone switched between 4G and Wi-Fi.
Before: 95%After: 99.9%
Faster delivery
About 7.5 weeks of setup was saved by reusing these components, and that time went into call quality and regional routing.
Provided tested sign-in and session handling for doctors and patients joining consultation rooms.
Tracked jitter, packet loss and session success by region, without collecting patient health information.
Gave the team a starting point for scaling on active consultation rooms instead of CPU load.
Kept the 12-region setup's spending visible as capacity grew and shrank through the day.
Trust and control
Designed to meet HIPAA, GDPR and PIPEDA requirements, with data residency controls for each region.
HIPAASignalling is encrypted end to end with rotating room keys, so the forwarding servers never see call content.
EncryptedCloud services autoscale across several regions, and traffic moves to another region automatically if one fails.
High availabilityThe client owns 100% of the IP in the custom scaling logic and monitoring dashboards.
100% ownedThe old platform sent every call through one large peer-to-peer signalling server. It crashed at about 5,000 concurrent users, usually in the 8 AM to 10 AM EST peak. With 2M+ patients, even a 1% failure rate meant thousands of missed appointments and frustrated specialists.
We replaced the single server with Selective Forwarding Units in 12 AWS regions, so each call takes the shortest network path. The cluster scales on the number of active consultation rooms rather than CPU load, adding capacity in under 30 seconds. New features reach 1% of traffic first.
Signalling is end-to-end encrypted with rotating room keys, so the forwarding servers never see call content. Every signalling event and log is scrubbed of patient health information before it goes into an audit trail. The platform was designed for HIPAA, GDPR and PIPEDA, with data residency controls.
If your telehealth platform struggles when clinics are busiest, we can help you plan a regional setup that scales with demand and keeps patient data private.