Your intake team stops re-keying.Patient data stays in your network.
Referral packets, discharge summaries and dictation, read on your own servers. Staff check what it isn't sure of, and you own the code.
No model call leaves your network.
Every field points to its source.
Models run on your hardware with outbound traffic denied by default. Each extracted value links back to the sentence it came from.
- 01
Documents arrive
Referral packets, discharge summaries, scanned pages and dictation, from the intake queues and record system they already land in.
- 02
Fields read and grounded
Diagnoses, medications, dosages and codes are extracted, and each one is tied to the source sentence a reviewer can click through to.
- 03
Unclear notes held back
When a note is ambiguous, the model abstains instead of guessing, and the field is left for a person to fill.
- 04
Staff verify
Intake staff open the referral with its fields pre-filled and check each one against its source sentence before it is used.
Every model call records who made it, for what purpose, which record it drew on, and which model version answered.
The same boundary, already running.
At a medical group and a private hospital.
Both run on hardware the client owns, with no outside AI service in the path. One reads clinical documents, the other scans.
Clinical documents a month
Read inside a specialty medical group's own network with zero external processors, at 0.95 field-level F1 on referral packets at cutover.
Read the case200,000Scans a month, moved on-premise
A Swiss private hospital moved its imaging AI off AWS onto its own servers, cutting ongoing costs by 95% and analysis time to 2-4 seconds.
Read the caseThirty minutes. You leave knowing whether the work is worth automating.
Three jobs it reads for your staff.
They keep the judgment and the sign-off.
None of it diagnoses or decides care. It takes the reading and the keying, and a clinician or coordinator signs off.
Referrals and discharge summaries
Read field by field: diagnoses, medications, dosages and codes, each with its source sentence. A note it can't read with confidence goes to a person.
Scans and dictation
Scanned pages read by OCR and dictation transcribed by a speech model tuned on your own recordings, on your servers, so every document becomes text.
A staff AI workspace
The consumer chat tools staff use on their own, replaced by one workspace on your servers, with SSO, role-scoped access and admin visibility.
See LumenThe questions compliance will ask.
Five answers you can test yourself.
Where patient data goes, who can open which record, what the audit trail holds, what trains on it, and who owns it.
No. It runs on your hardware or private cloud, calls no outside AI service and sends no telemetry, so there is no new processor or subprocessor to assess.
Not through this system. Access is checked at retrieval, before the model sees anything, so a record the user may not open never enters the prompt.
Append-only rows: who asked, for what purpose, the source record, the model version, and hashes of prompt and response, kept as long as your retention rules require.
No. Training or fine-tuning on production clinical data stays off unless you approve a separate lane, documented with its own data flow diagram.
You own the code, models and documentation on delivery. It all runs on your hardware, so your team or any auditor can inspect it, down to the model weights.
Start with one document type.
Or with a two-week review.
One workload at a fixed scope and price, accepted on your own documents. Your team gets the runbooks and runs it after handover.
Automate one workload
Referral intake or discharge summaries, run beside your current process. Cutover comes when it meets an agreed threshold on documents your staff labelled.
A two-week review
We go through your intake and records work and tell you which document flows are worth automating, and which are not.
“Bring the intake queue your staff spend the most time on. On the first call we walk through what the system would pull from those documents, what it would hold back for a person, and whether it is worth building.”
Sebastian Mondragon, Founder of Particula Tech