Braynex, our work
MedChronoLegal Tech / Healthcare AI2026

Six hundred pages in.
One defensible timeline out.

We built MedChrono, an AI processor that reads hundreds of pages of messy medical records, digital or scanned, and builds a clean, dated chronology of every diagnosis, treatment and visit in minutes.

What we didAI Engineering, Full-stack Development, OCR & Document Pipeline, Prompt Engineering
StatusIn private beta with US personal-injury firms
Scroll, and watch the records get clean
01The problem

The records do not agree with each other.

In injury and malpractice cases, a law firm has to produce a medical chronology: a date-ordered account of everything that happened to a patient across every provider. Building one by hand means reading stacks of PDFs, scans and Word docs from different hospitals, pulling date, provider and event out of inconsistent formats, and removing the same visit reported five different ways. It is slow, expensive, and it becomes legal evidence, so it has to be right.

MedChrono does that work automatically. Records go in through three lanes shaped around how a real case is organized, and come back as a structured, dated chronology, each entry broken into diagnosis, symptoms, treatment, notes and prognosis. It reads clean digital PDFs and low-quality scans alike, and every row traces back to the file it came from.

Three of the six hundred and eleven pages, exactly as they arrive. Zoom in on the last one: section 2 is the date of referral, and the line is blank.

One injury, five documentsNot one of them matches
Northgate, ERLumbar strain02/03/19
Dr. M. AyresBack injury2 Mar 19
Discharge summaryLumbar sprain03/02/2019
FairhavenL4-L5 disc bulge18-04-2019
PhysiotherapyLower back painno date

The first three are the same day, written three ways. Two of them could be the 2nd of March or the 3rd of February, and nothing on the page says which. A paralegal resolves that by hand, 611 pages at a time.

02The pipeline

Five stages. Two of them think.

1Read
2Chunk
3ExtractAI agent
4De-duplicateAI agent
5Order
ER admissioner-admission_northgate_2019-03-02.tiff2019-03-02Lumbar strain, Dr. M. Ayres
MRI reportMRI-report_Fairhaven_2019-04-18.pdf2019-04-18MRI confirms injury, Fairhaven
Physio referralphysio-referral_undated.docxno datePhysiotherapy referral
Consult notesayres-consult-notes_p4.jpg2019-06-11Follow-up consult, Dr. M. Ayres
Discharge summarydischarge-summary_2020-01-09.pdf2020-01-09Recovered and discharged
Hybrid ingestion

Digital PDFs are read straight out of the file. Scans and photos go through OCR. PDFs that are secretly just pictures of pages get caught and rerouted.

03The cost of one row

One line in the timeline.
Six hundred and eleven pages underneath it.

Not hidden, just heavy. Every one of those pages was read, and this row still traces back to the single page it came from.

What the legal team sees
2019-03-02Lumbar strain, Dr. M. Ayresayres-consult-notes.pdf · p.4

611 pages were read to produce that one row. It came from exactly one of them, and it still says which.

What it took to put it there
0pages read
0providers reconciled
0formats, incl. scans
0model passes
0source page cited
04The rule

Three proofs, or the event
does not exist.

A matching engine can afford to rank a near-miss. A chronology cannot: it becomes evidence. So MedChrono does not weigh a maybe, it refuses it.

An explicit diagnosisWritten in the record, never inferred from context.
and
An exact dateA real calendar date, not “the following spring”.
and
A named treaterThe doctor or the hospital, named on the page.
ayres-consult-notes.pdf · page 4
The raw text, as it came off the page
…pt seen 02/03/19 by Dr. M. Ayres, c/o lower back pain following MVA. Dx: lumbar strain. Rx: physio 6/52, NSAIDs prn. Px: good…
Diagnosis Date TreaterAccepted
What enters the chronology
2019-03-02Lumbar strain, Dr. M. AyresSourced

Broken into five clinical fields: diagnosis, symptoms, treatment, notes and prognosis. Every one of them lifted from the page, none of them inferred.

physio-referral.docx · undated
The raw text, as it came off the page
…referral to physiotherapy. Dx: lumbar strain. Referring clinician: Dr. M. Ayres. 6 sessions authorised. [date of referral not stated]
Diagnosis Date TreaterRefused
What reaches the reviewer instead
no datePhysiotherapy referral, Dr. M. AyresNeeds a human

Two of the three surrounding pages are dated March 2019. MedChrono could infer a date from them and be right most of the time. It does not. A guessed date in a chronology is not a bug, it is a false statement in evidence.

It never invents a dateIf the record does not say when something happened, the row is flagged, not guessed.
It never names a treater it did not readAn unattributed event stays unattributed, rather than being pinned on the nearest doctor on the page.
It never merges what it cannot prove is the sameTwo events collapse into one only on a matching date, treater and diagnosis.
05The interface

Twenty-two events.
Two it will not place.

Two hundred and seventy pages, read and ordered, every row traceable to the page it came from. And two events with no date anywhere on the record, which it declines to invent one for.

06The engagement

The hard part was never the model.

Any system can produce an answer. The engineering is in knowing when not to. Across six hundred pages of contradictory paperwork, that discipline is the difference between a summary and evidence.

Constraints before capabilityWe define what a system must never do before we decide what it should. In regulated work, the boundary is the product.
Built to be auditedSomeone will check the output. Every claim a system makes traces to a source, so verification takes seconds and not trust.
A scope both sides can seeMedChrono's team held the product and the legal domain. We held the engineering. Neither side had to guess where the line was.

Work where being wrong is expensive.

That is what we build for.

Talk to us