A live ambient scribe for clinical visits
How a doctor-patient conversation becomes a reviewed clinical note, on the HIPAA-compliant platform I lead at Atelia.
Details are kept general to protect the platform and its patients.
Notes are the bottleneck
Writing clinical notes pulls a clinician's attention away from the patient. An ambient scribe listens to the visit and drafts the note as it happens, so the clinician reviews a draft instead of writing from scratch.
In healthcare the bar is higher than a good demo. The output has to be right, private and traceable before it can go anywhere near a patient's record.
From visit audio to the record
- 01audio
- 02ws stream
- 0330+ agents
- 04clinician review
- 05record
01
audio
The visit's audio is captured while the doctor and patient talk.
02
ws stream
Audio streams to the backend over native WebSockets, so the system works on the conversation while it is still happening and can surface clinical suggestions during the session.
03
30+ agents
Specialized agents with 200+ tools handle documentation, summaries and clinical workflows. A router sends each step to the right agent, and retrieval uses embeddings and pgvector in PostgreSQL. Heavier work such as summarization and document generation runs in background workers.
04
clinician review
Drafts and suggestions wait for a clinician. Human review is a step in the workflow, not an afterthought.
05
record
Only reviewed content is written to the clinical record.
The two hardest calls
Checkpointed state
Agent workflows here are multi-step and route work between specialized agents. They are orchestrated with LangGraph, and state is checkpointed between steps.
That costs extra storage and more care around state shape. In return a failed step can be retried from where it stopped instead of from the start, and every intermediate state can be inspected when something looks wrong. Structured logging, retries and graceful degradation sit around it.
Clinician review before the record
The fastest path would be writing AI output straight into the chart. Instead, every draft waits for a clinician.
It adds a step to every visit. In a clinical record a wrong note costs far more than a slow one, and review keeps a person accountable for what the record says.
In production
- Used daily in production by doctors, hospitals and patients on a HIPAA-compliant platform.
- Real-time suggestions during live sessions reduced latency and helped clinicians make faster decisions.
- Multi-step agent reasoning increased the relevance of generated clinical suggestions.
- Part of an AI layer I lead that now runs 30+ agents and 200+ tools.