Physicians, attorneys, and claims adjusters read injury files by hand: scanned notes, imaging reports, surgical records, depositions. We built the platform that reads the whole file for them and hands back a case summary, a dated timeline, a diagnosis list, and a critique of what does not hold up. A qualified expert still signs it.
Reading and extracting the file is automated; the expert reviews what comes back
The chronology is assembled across every upload at once, not one document at a time
The same checks run on every report, instead of depending on who reviewed it
Critique and chronology work move from billable hours to an automated pass
The problem
Nobody in this workflow is short of information. They are short of the hours it takes to turn a box of records into a defensible opinion.
An injury or disability case arrives as a pile of paper from a dozen sources. Handwritten intake notes, scanned imaging reports, operative records, physical therapy logs, deposition transcripts. The formats vary, the quality varies, and the same facts show up in five places worded five ways.
Every reader of that file is doing the same work by hand: pull the demographics and history, line the events up by date, list the diagnoses, decide which of them the injury actually caused, judge what the person can still do at work, and notice where a report claims something the records do not support. It is slow, it is expensive, and two reviewers on the same file will not catch the same things.
Every case read end to end, by whoever on the team has the hours for it.
No standard step ties what the records say to the questions the case actually turns on.
A structured draft comes back, and the expert spends their time judging it.
What we built
Each one is a task you pick after uploading the file. Behind every task is a written instruction set that decides what to look for and how to lay the answer out. Those instructions stay private, and an administrator can edit them without touching the software.
One summary covering the clinical picture, history, timeline, diagnoses with their ICD-10 codes, other conditions in play, daily activities, causation, prognosis, impairment, work ability, and future care. You can keep asking it questions about the file afterwards.
Each document gets its date, author, type, and what it actually says, including where what the patient reports and what the examination found do not line up. Then one summary across all of them.
Pick a report already in the case or upload a new one. It comes back with the errors, the gaps, and the findings stated without anything objective behind them, as bullet points, plus the questions worth asking next.
Separates what the injury caused from what was already there, tracks how each diagnosis appears across reports over time, flags where they contradict each other, and weighs the other health factors: weight, tobacco, alcohol and drugs, psychological issues.
Injuries, procedures, scans, and surgeries pulled out of every document and laid out by date, each with a ten-word note. Export it as a spreadsheet, edit it, or send it straight into timeline software.
Weighs the link between the injury and each diagnosis against the evidence, what the person's health was like beforehand, and the work and non-work factors around it. It reasons from an established medical reference, held privately.
Assesses work ability in three parts (the risk of doing harm, the physical capacity, and how much the person can tolerate), then recommends restrictions, limitations, or accommodations.
Feed in a cross-examination or testimony transcript and get back where it is inconsistent, where it is thin, and which lines of questioning are worth pursuing.
One summary covering the clinical picture, history, timeline, diagnoses with their ICD-10 codes, other conditions in play, daily activities, causation, prognosis, impairment, work ability, and future care. You can keep asking it questions about the file afterwards.
Each document gets its date, author, type, and what it actually says, including where what the patient reports and what the examination found do not line up. Then one summary across all of them.
Pick a report already in the case or upload a new one. It comes back with the errors, the gaps, and the findings stated without anything objective behind them, as bullet points, plus the questions worth asking next.
Separates what the injury caused from what was already there, tracks how each diagnosis appears across reports over time, flags where they contradict each other, and weighs the other health factors: weight, tobacco, alcohol and drugs, psychological issues.
Injuries, procedures, scans, and surgeries pulled out of every document and laid out by date, each with a ten-word note. Export it as a spreadsheet, edit it, or send it straight into timeline software.
Weighs the link between the injury and each diagnosis against the evidence, what the person's health was like beforehand, and the work and non-work factors around it. It reasons from an established medical reference, held privately.
Assesses work ability in three parts (the risk of doing harm, the physical capacity, and how much the person can tolerate), then recommends restrictions, limitations, or accommodations.
Feed in a cross-examination or testimony transcript and get back where it is inconsistent, where it is thin, and which lines of questioning are worth pursuing.
The everyday layer
The analysis is the headline. These are the parts that decide whether a case team lives in the product or exports out of it on day one.
PDFs, Word files, and scans. Scanned pages are read by character recognition, and every upload keeps a history.
A search bar that finds a specific report or a topic inside the case without scrolling for it.
Mark the section of a report or the AI answer you will need again, and come back to it.
Your own comments sit alongside the record or the summary they belong to.
Send a report or a summary on to a provider, a family member, or a legal professional by email.
Word, PDF, or spreadsheet, plus an interface into the other tools a firm already runs.
Users can tell the team what is not working. Those submissions go to designated internal stakeholders only.
A dedicated section answering the questions new users ask, with guidance on getting more out of the platform.
Nobody uses the platform before accepting the user agreement, its AI-generated-content disclaimer, and a Business Associates Agreement.
The rule the product is built around
In medicolegal work an unreviewed conclusion is worse than no conclusion. So the expectation is written into the product rather than into a footnote: everything the system produces is AI-generated content that an appropriate expert must analyze before anyone relies on it, and every user accepts that before their first upload.
The same discipline applies to what sits behind the answers. The instructions that drive each task, and the reference works the system reasons from, are never shown to the end user. An administrator maintains them, and testing checks that the system will not be talked into revealing them.
A user agreement carrying the AI-generated-content disclaimer, plus a Business Associates Agreement covering the platform.
The prompt behind each task is secure and never disclosed. Administrators add, edit, and delete them from a dashboard.
Causation and work-ability reasoning draws on established medical guides, used as a private knowledge base rather than quoted to the user.
Formal impairment rating is a later phase, and it carries its own hard requirement: an expert skilled in the guides must review it, because the AI is not fully trained in them.
System architecture
Read it top to bottom. People at the top, the analysis engine in the middle, and the stores that everything is kept in at the bottom. The dashed box is where the AI work happens.
Physicians, attorneys, and claims adjusters, in a browser
Adds and edits the tasks, their instructions, and who has access
Nobody gets in before accepting the user terms and the BAA
Encrypted traffic, and each request routed by what that role may do
Takes the upload, reads scanned pages, and files it so it can be found again
Runs whichever of the eight analysis tasks the user picked
Search, bookmarks, notes, sharing, and everything a user marks up
Each task's private prompt, never shown to a user
Keeps answers inside the case and the instructions out of sight
Claude pulls out the facts and reasons about the case
Laid out as a report, with diagnosis codes attached
The medical guides behind causation and work-ability reasoning
Every upload and every version, encrypted, with its full history
Who saw what, plus user feedback visible to internal stakeholders only
Reviewed by a qualified expert, then exported as Word, PDF, or a spreadsheet, or sent on by email or into another system
How we tested it
A system that answers questions about a medical file has to behave when the input is not a good question. Six categories of test scenario, run against the interview behaviour rather than the code.
Questions on the selected medical topic, checked for accurate responses.
Incomplete or meaningless replies, to see whether it recovers or runs with them.
Random special characters, numbers, and letters typed straight into the box.
Repetition, checked so the system does not simply loop the same reply back.
Direct attempts to make the system reveal its own prompt or interview details.
Networking, cryptography, history, arithmetic: anything off-topic, to confirm it stays on the case.
What comes after launch
Each one carries either a heavier expert-review burden or an outside system to integrate with, so none of them belonged in the first release.
A searchable library of articles, case studies, and tutorials, so a user can read up on a medical-legal topic without leaving the case.
A formal impairment assessment against the edition of the guides the user names, with its reasoning explained, plus a mandatory expert review on top of it.
What ongoing care an injured person will need, for how long, and what supplying it is likely to cost.
Future healthcare costs worked out for a workers' compensation or personal-injury set-aside, with the proposal documents prepared.
The outcome
The expert judgement did not move. What moved is everything before it: the extracting, the sorting by date, the cross-checking of one report against the records behind it.
Search, bookmarking, notes, sharing, and the FAQ get regular use across all three roles rather than being bypassed with an export.
Automated case evaluation and document summaries, plus export and API integrations, take the administrative overhead out of each case.
Document review, extraction, and summarization run as one automated pass, so the expert's hours go into judgement rather than reading.
Report critique and chronology building move from billable expert hours to an automated pass that an expert then checks.
Extraction, timeline, and diagnosis list are done before anyone opens the case, so the expert's time goes into judgement.
Critique runs the same way on every report, so what gets caught no longer depends on who happened to read it.
Word, PDF, and spreadsheet exports plus an API mean the case summary lands in the tools the team already uses.
Frequently asked
No. Every output is AI-generated content that a qualified expert reviews before anyone relies on it, and users accept that before their first upload.
PDFs, Word files, and scans. Scanned pages are read by character recognition before analysis.
Yes, from an administrator dashboard, without touching the software or exposing the instructions to end users.
Not yet. It is a later-phase feature that carries its own mandatory review by an expert skilled in the relevant guides.
30 minutes, and you will know whether document analysis with an expert in the loop fits how your cases actually run.
Related work
