Case Study

Supercharging the personal injury space

We created the first platform, powered by responsible AI, to assist doctors testifying in legal cases. Our system streamlines the review of tens of thousands of pages, pinpointing critical information to help doctors form well-informed opinions in hours instead of days. Designed with transparency in mind, it allows doctors to easily fact-check AI-generated insights, ensuring they can confidently validate and take ownership of their findings.

Visit Website

Problem statement

Evaluating personal injury cases costs $20-50K per review of 20-30 hours, diverting valuable doctor time from helping their patients to sifting through thousands of disorganized, low-quality documents.

Our Goal

Create the evaluation for the doctor within 10 minutes, allowing the doctor to come to a strong conclusion in a matter of hours.

Document ingestion

The problem we’re addressing: large PDFs are a mix of various document types—doctor assessments, invoices, insurance policies—many of which are either of low significance or have highly specific relevance. Sending these mixed documents to an LLM can degrade the quality of the results and makes citations and fact-checking impossible.

Our approach

We use an algorithmic approach combined with a vision model we built and trained. This allows us to understand where one document ends and another begins. Our current accuracy stands at 97%, well above the required threshold. We can OCR and split 20 PDFs, totaling 2,000 pages, in under 10 minutes.

Integration with the rest of the process

We use a custom visual model trained to detect document boundaries within PDFs, creating a JSON file that maps OCRed content in our database. The result? We transform chaotic, disorganized data into a structured format, ready for seamless analysis—so our clients never have to worry about organizing data again.

Extracting value from documents

The challenge we tackle: navigating thousands of pages filled with low-value information to uncover the critical high-value insights hidden within.

Model Evaluation

We created our own layer of logic to quickly test and deploy “agents” that are customized for a specific task (e.g., looking for events, or answering nuanced medical questions based on dense text). We rely on general purpose models and custom trained models, depending on the context and success rate.

API Calls are easy, consistent results are not

We iterate on results quickly, create our own evals frameworks to inform prompt engineering, and we iterate on model selection.

Sample results

Surgeon Report

____ presented to Emergency Department after motor vehicle collision. is reported that she was the restrained driver of a vehicle that was struck from behind. She arrived to ED via ambulance with the chief complaint of neck pain and low back pain. Exam: Diffuse tenderness to palpation mid-cervical spine. Tenderness to palpation in lower lumbar, sacral area. Normal neurological exam. 

Cervical spine radiographs interpreted by ____. Dr. ____ summarized: Alignment is normal. No fracture or subluxation. Disc spaces well maintained. Prevertebral soft tissues unremarkable. Lumbar spine radiographs interpreted____. Dr. ____ summarized: Alignment is normal. No fracture or subluxation. Disc spaces well maintained. Minimal marginal osteophytes at several levels.

Gemini

____, a __-year-old female, arrived by ambulance at the ____ Hospital emergency department at 1:47 AM following a motor vehicle collision. She complained of moderate neck and lower back pain. ____ ____ reported that the collision was a rear impact, she was the restrained driver, and had not walked since the accident. The physician noted no bleeding. Her vital signs were: temperature 97.6 degrees F, pulse 105 bpm, respirations 18, and blood pressure 175/103 mmHg. A physical examination revealed diffuse paravertebral muscle tenderness in the mid-cervical level and tenderness in the lower lumbar and sacral levels of the back. No bruising or swelling was noted. Radiology reports indicated no fractures identified in the cervical or lumbar spine. _ ____ was diagnosed with neck and back pain/strain. She was prescribed Flexeril 10 mg one tablet three times daily as needed for spasm, and Ultram 50 mg one tablet every eight hours as needed for pain. ____ was provided educational materials about motor vehicle collisions, general precautions, neck sprain/strain, and back and neck pain. She was discharged and instructed to follow up with ____.

MedLM

The patient was involved in a motor vehicle collision and presented to the emergency department with neck and back pain. She was restrained and the collision was a rear impact. She had no loss of consciousness.

On physical examination, the patient had diffuse paravertebral muscle tenderness in the mid cervical level and tenderness in the lower lumbar, sacral level. There was no bruising or swelling.

Radiographs of the cervical and lumbar spine were negative for fracture or subluxation.

The patient was diagnosed with neck and back pain/strain and was discharged home with instructions to follow up with her primary care physician within 1 week.

She was prescribed Flexeril 10 mg oral tablet 3 times daily and Ultram 50 mg oral tablet every 8 hours as needed for pain.

Web app integration

We provide a seamless web application experience by managing every detail: from user experience research and interface design to ensuring authentication and authorization layers meet client requirements. In this specific project, we also ensured full compliance with HIPAA and ISO standards.

Web application

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi.