Agent skill · software engineering · jeremylongshore
abridge-performance-tuning
Optimize Abridge clinical AI integration performance for high-volume
Why this skill is useful
Adds executable scripts for optimizing audio streaming and note generation pipelines specific to Abridge's clinical AI integration.
What it needs
Requires npm installed locally. About 4k tokens when loaded. Last updated 2026-08-07. 2,604 stars on the source repository.
What this skill does
Abridge Performance Tuning Overview Performance optimization for high-volume Abridge deployments. Large health systems process thousands of encounters daily — latency in note generation directly impacts clinical workflow throughput. Performance Targets Metric Target Critical Threshold -------- -------- -------------------- Audio stream → first transcript < 2s > 5s Encounter → completed note < 30s > 60s Note → EHR push < 3s > 10s Patient summary generation < 10s > 30s Concurrent sessions per org 100+ < 50 Prerequisites A staging or sandbox environment with synthetic encounter fixtures and a representative, authorized load profile. Baseline latency, error-rate, and concurrency measurements for the current release, with timestamps and measurement window recorded. Clinical, EHR, and security owners agree on safe concurrency limits and a rollback threshold before changing production traffic behavior. Instructions Step 1: Audio Streaming Optimization Step 2: Note Generation Pipeline Optimization Step 3: Connection Pooling for FHIR Push Step 4: Performance Monitoring Dashboard Output Optimized audio streaming with 100ms chunking Adaptive polling for note generation (500ms → 3s backoff) Connection-pooled FHIR batch pushes Real-time performance monitoring with P50/P99 latency tracking Examples Run a staged load test using synthetic cardiology encounters at the approved concurrency level. Capture baseline and tuned P50/P99 note latency, FHIR push success rate, and audio-drop rate over the same measurement window. Enable connection pooling and adaptive polling only after the baseline is saved, then compare the two snapshots without logging transcripts or identifiers. If P99 latency, error rate, or downstream EHR saturation crosses the agreed threshold, disable the change and return traffic to the previous configuration before investigating the aggregate metrics. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills jeremylongshore/abridge-performance-tuning