The challenge
What was at stake
A large multi-specialty hospital network managing over 2,800 daily outpatient visits across 12 departments faced severe pre-assessment bottlenecks — average OP wait times had climbed to 64 minutes, with 44% of physician consultation slots consumed by repetitive intake questioning.
The hospital's internal audit identified that 68% of no-show appointments were linked to a broken pre-screening process. Due to strict HIPAA mandates, the hospital could not route any patient records to external cloud-hosted LLMs, with potential compliance penalties exceeding $4.5M per breach incident.
The approach
What we built
We engineered a phased, fully self-hosted clinical AI system deployed entirely on the hospital's existing on-premise server infrastructure. Phase 1 introduced a low-latency voice triage agent that conducted structured pre-assessment interviews.
Phase 2 expanded this into a full multi-agent clinical swarm: a private quantized language model cluster (running on 8x NVIDIA A100 80GB nodes) ingested and cross-referenced each patient's 10-year EMR history, flagging 42 clinical risk indicators and generating a structured pre-consultation brief. The system processed an average of 2,400 patient records nightly with zero external data transmission.
Inside the system
Key capabilities
Low-Latency Voice Triage Agent
Outbound pre-assessment calls averaging 3.8 minutes in duration, covering 24 standardized clinical intake fields with a 96.2% patient completion rate.
Fully Self-Hosted PHI Isolation
All model inference runs on private GPU nodes inside the hospital's local network. Zero bytes of patient data leave the firewall.
Predictive EMR Intelligence
The swarm cross-references 10 years of historical EMR data to generate a ranked differential risk brief with per-diagnosis confidence scores (avg. 91–96%) delivered to physicians 60 minutes before consultation.
Interactive Doctor Consultation Portal
Authenticated physicians access a secure web portal to voice-query the clinical agent during live consultations, receiving contextual summaries within 1.2 seconds.
Under the hood
Technology
- Self-hosted quantized Mixtral-8x7B and Llama-3-70B-Instruct clusters
- LiveKit voice gateway
- private FHIR-compliant vector store
- HIPAA-audited local inference server
The outcome
What changed
Average OP pre-assessment time reduced from 64 minutes to 12 minutes (81% reduction). Physician intake workload cut by 74%. Appointment no-show rate dropped from 28% to 6% within 180 days of go-live.
- 81%
- Wait time reduction
- 0
- Phi egress events
- 2,400+
- Emr records/night