AutoTox: AI-Powered Toxicology Interpretation, Deployed at UTMB

Most clinical AI never leaves the demo. AutoTox is in the daily workflow at a hospital — drafting toxicology interpretations that a pathologist verifies and signs out.

I co-led engineering on AutoTox with Gladson John, under Dr. Peter McCaffrey (Director of Pathology Informatics at the John Sealy Diagnostic Center; Chief AI Officer, UTMB). It went from idea to production in five weeks — a timeline McCaffrey called “unheard of.” The details below are drawn from UTMB’s Retool case study.

The problem: a slow, manual sign-out

Toxicology sign-out is a reading task that doesn’t scale. For every case, a pathologist works through more than 40 quantitative metrics alongside the patient’s demographics and medication list, then writes an interpretation — roughly 15 minutes per case.

UTMB’s four-person pathology team could clear only about 50 cases a week that way, and demand kept climbing. The bottleneck wasn’t judgment; it was the time it takes a human to read 40 numbers, over and over, without missing one.

What we built

AutoTox is a decision-support tool wired into the live clinical pathology workflow. It pulls each case’s results, asks a large language model to draft a plain-language interpretation, and hands the pathologist a single screen to review, correct, and sign out — without any patient data leaving UTMB’s governed environment.

How a case flows through it

Extract. A Python job runs a SQL query that pulls the case’s toxicology results out of UTMB’s medical record system into a PostgreSQL database.
Interpret. AutoTox sends those results to OpenAI GPT models through a private, HIPAA-compliant Azure backend, which returns a simplified report summary in seconds.
Review. The pathologist sees a case list with metadata, the raw test results, and the AI’s interpretive text side by side — then verifies the findings and leaves feedback on the AI output.
Sign out. Verified results flow back into the medical record.

  Urine tox panel
         │
         ▼
  Python + SQL  ───►  PostgreSQL          ┐
         │                                │  all inside UTMB's
         ▼                                │  security perimeter
  OpenAI GPT  (private, HIPAA Azure)      │  (self-hosted Retool)
         │                                │
         ▼                                │
  Plain-language summary  (seconds)       ┘
         │
         ▼
  Pathologist reviews + verifies  ───►  Medical record

The AutoTox pipeline, as described in UTMB’s Retool case study. The model drafts; a pathologist decides.

The human stays in the loop

Every AI-generated report is verified and validated by a pathologist before it is signed out. The model never has the last word — it produces a first draft that a licensed physician confirms or corrects.

That design also turns the AI’s stamina into a safety feature. Because it reads all 40-plus levels on every single case, AutoTox surfaces errors a human skimming that many numbers might miss — so the tool didn’t just speed sign-out up, it improved accuracy.

The results

Metric	Before	With AutoTox
Sign-out time per case	~15 min	~5 min (66% faster)
Cases reviewed per week	~50	~500 (10×)
Toxicology metrics per case	40+ read by hand	40+ read automatically, every case
Idea to production	—	5 weeks

“It looks professional, it’s easy to use, it’s interactive, it doesn’t go down.”— End-user feedback, from the Retool case study

Built to run inside a hospital

Getting AI into a clinical workflow is mostly a governance problem, not a modeling one. AutoTox runs on a self-hosted Retool front end with a Python and PostgreSQL data layer, and reaches OpenAI GPT models only through a private Azure backend.

Self-hosting is what made that acceptable: it kept the whole system inside UTMB’s security perimeter and compliant with both HIPAA and internal protocols. Protected health information never touches an un-governed endpoint — the same boundary condition that drives my dissertation research on privacy-preserving clinical LLMs.

Why it matters

AutoTox is the kind of clinical AI I care about building: not a proof of concept, but a tool clinicians actually use, that saves real time, and that keeps a physician accountable for every result. The pattern — extract structured data, let a model draft the tedious interpretive layer, and put an expert in front of the output — generalizes to a lot of high-volume interpretive sign-outs beyond toxicology.

The team

Peter McCaffrey, MD — Director of Pathology Informatics, John Sealy Diagnostic Center; Chief AI Officer, UTMB
Gladson John — lead engineer
James Weatherhead — lead engineer

Case study: How UTMB Improved Diagnostic Testing for Thousands of Patients Using AI (Retool, 2024)
Demo & discussion: AutoTox on YouTube

Facts and figures on this page are drawn from UTMB’s Retool case study, “How the University of Texas Medical Branch Improved Diagnostic Testing for Thousands of Patients Using AI” (2024).

AutoTox is a clinical decision-support tool. Every interpretation is reviewed and verified by a licensed pathologist before sign-out; the system does not make autonomous diagnostic decisions.