Your imaging AI moves in-house.Patient scans stay in your network.
Imaging inference, diagnostics triage and clinical documents, run on your own servers. The read stays with the radiologist. You own the code.
The scan never leaves your network.
The radiologist still makes the read.
Models on your own GPUs, wired into PACS and RIS, calling no outside AI service. Results return to the worklist radiologists use.
- 01
Studies arrive
CT, MRI, X-ray and ultrasound studies, sent from PACS as each scan completes. Referrals and discharge summaries from the queues they already land in.
- 02
Scored on your GPUs
Each study gets a probability per finding, not a diagnosis. Clinical documents are read field by field, each value tied to its source sentence.
- 03
Radiologist review
Likely findings rise to the top of the worklist with the flag attached. A radiologist reads the study and makes the call.
- 04
Back in your worklist
Priority flags and preliminary findings land in the PACS and RIS worklists you already run. Radiologists' screens stay the same.
Every inference is logged with its source study or record and the model version, and every document field links to the sentence it came from.
Imaging and clinical text, in production.
At a Swiss hospital and a medical group.
Both run on hardware the client owns, with no outside AI service. The Chilean model was scored on records its own staff labelled.
Lower ongoing infrastructure cost
200,000 scans a month moved from AWS to GPUs in the hospital, with results in 2-4 seconds instead of 8-12. About $5.4M less than three more years on AWS.
Read the case45,000Clinical documents a month
Read inside a Chilean specialty medical group's own network with zero external processors. 0.95 field-level F1 on referral packets, a source span on every field.
Read the caseThirty minutes. You leave knowing whether the work is worth automating.
Three jobs that queue up for clinicians.
It scores and sorts, clinicians decide.
None of it makes a diagnosis. It scores studies and extracts fields, and a clinician decides on every finding.
Imaging AI, moved in-house
Inference already running in a cloud account, moved onto GPUs in your building. Both paths run in parallel while the load shifts to local servers.
Diagnostics triage
Studies scored for likely findings as each scan completes, so likely urgent ones reach a radiologist first. The read stays with the radiologist.
Clinical documents
Referrals, discharge summaries and dictation read into structured fields, each linked to its source sentence. Unclear notes go to a person.
What your IT and privacy teams will ask.
Five answers you can check yourself.
Where patient images go, how access is recorded, what changes in PACS, how cutover works, and who owns the result.
No. Inference runs on your servers or your private cloud and calls no outside AI service. In the Swiss build, scans never left the hospital building.
In the Chilean build, yes: every inference logs the user, purpose of use, source record and model version, in append-only tables the compliance team queries.
PACS sends studies to local inference servers instead of a cloud endpoint, and results return to the worklist radiologists already use. Their screens stay the same.
Old and new paths run in parallel on live studies while load shifts over. The Swiss cutover ran on a low-volume weekend, with the cloud path kept ready as fallback.
You do. Code, models and documentation are yours on delivery and run on your hardware, so your team or any auditor can inspect all of it.
Start with one workload.
Or with a two-week review.
Fixed scope, fixed price, and the code is yours at the end. We hand over runbooks and stay until your team runs it day to day.
Move or build one workload
One modality or one document queue, measured in turnaround and running cost. Old and new run side by side until the switch is safe.
A two-week review
We go through your imaging and document workloads, cloud bills included, and tell you which are worth running in-house and which are not.
“Bring your study volume and your current inference bill. On the first call we work out whether running it in-house pays off, and if it doesn't, we say so.”
Sebastian Mondragon, Founder of Particula Tech