A Surgeon Built His Own Tools with Codex: How AI Is Reshaping Medical Literature Research

When a Surgeon Starts Writing Code

The intersection of medical professionals and AI coding tools is giving rise to a surprising range of applications. The case shared by surgeon @pridgenmd illustrates an unconventional user profile: a doctor without a professional software engineering background who used OpenAI's Codex to build his own research tools and dramatically improve how he searches and organizes scientific literature on PubMed.

The value here isn't in the technical complexity of what was built — it's in the trend it reveals. AI coding assistants are putting the ability to "build your own tools" directly into the hands of domain experts. In the past, a physician who wanted a custom literature screening system would have had to rely on an engineering team or commercial software. Today, he can describe what he needs in plain language and let Codex generate working code.

Extracting Signal from a Sea of Literature

One of the biggest pain points in medical research is sheer volume. PubMed indexes tens of millions of biomedical articles. For a clinician trying to track the latest developments in a given field within a limited amount of time, reading papers one by one is simply not feasible.

@pridgenmd mentioned that one of his earliest tools was built to "pick out information from a lot of articles." At its core, this kind of tool offloads the mechanical work of information extraction to an automated program, freeing the physician to focus on tasks that genuinely require professional judgment — like evaluating the rigor of a study's methodology or assessing the clinical relevance of its conclusions.

This "personalized research assistant" approach is fundamentally different from general-purpose reference management software. It's built around a specific user's workflow, precisely matching one surgeon's search habits and areas of focus.

PubMed is a free biomedical literature database maintained by the U.S. National Library of Medicine (NLM), currently indexing over 37 million article abstracts and full-text links across medicine, nursing, pharmacy, dentistry, and related fields. While PubMed offers features like Boolean logic search and MeSH subject headings, extracting and comparing specific data fields — such as sample size, intervention type, or outcome measures — across large sets of papers on niche or cross-disciplinary topics still requires significant manual effort. This is precisely where automation scripts deliver the most value: by calling PubMed's E-utilities API, a program can batch-download structured metadata and abstracts, then use natural language processing to extract and categorize target fields — compressing what might take hours of manual screening into minutes.

The Unexpected Payoff of "Explaining"

One particularly thought-provoking detail in this case is @pridgenmd's observation about Codex: what surprised him was that the act of explaining his needs to the AI itself "unlocks so much more context about what I want and what I want to do."

This points to a frequently overlooked value in AI-assisted programming. When users are forced to articulate clearly what they want to an AI, they are effectively clarifying and organizing their own requirements. That process often helps surface vague ideas and sharpen real goals. In other words, the conversation with AI isn't just about getting a machine to understand a human — it also helps humans better understand themselves.

For professionals without a technical background, this is especially significant. They don't need to know programming syntax — they just need to be able to clearly explain their domain-specific needs, and the tool handles the rest.

Code Security Covered Too

@pridgenmd also mentioned using Codex to build a "security plug-in," and was impressed by the granularity of what it caught — "The things it catches are just incredible in the detail."

This deserves separate attention. For a non-professional developer, self-written code can harbor all kinds of vulnerabilities and risks that the author simply lacks the expertise to detect. AI tools play the role of "code reviewer" here, helping amateur developers avoid pitfalls they wouldn't have noticed on their own. In contexts involving sensitive medical data, this is especially critical — security isn't a nice-to-have; it's a hard requirement.

Code security and compliance requirements in healthcare contexts are far stricter than in general applications. In the United States, software handling protected health information (PHI) must comply with HIPAA's technical safeguard provisions, including specific requirements around data encryption, access controls, and audit logging. In the EU, GDPR and the Medical Device Regulation (MDR) apply. For clinicians building their own tools — even tools used solely for personal research with no direct patient data involvement — common issues like improper API key management, vulnerable dependency versions, and insecure local storage can still create data breach risks. AI tools can identify typical problems at the static code analysis level, such as hardcoded credentials, SQL injection risks, and insecure HTTP requests, effectively providing an automated first-pass review for amateur developers. However, this is not a substitute for formal security audits.

What This Case Tells Us

Putting the pieces together, we're seeing a new paradigm take shape: domain expert + AI coding tool = self-sufficient, customized productivity.

Physicians are no longer passive consumers of software. They can now rapidly build fit-for-purpose tools grounded in real clinical and research needs — becoming "citizen developers" in their own right. Behind this shift is AI dramatically lowering the barrier to programming, shrinking the gap between specialized knowledge and technical implementation.

That said, some grounded perspective is warranted. Single Twitter case studies tend to highlight success stories. In practice, the reliability of AI-generated code, its suitability in healthcare compliance environments, and the cost of validating its outputs are real considerations that can't be ignored. In medicine especially, all tool outputs require rigorous oversight from trained professionals. AI improves efficiency — it doesn't replace professional judgment.

But regardless, this case offers a concrete and vivid example: the true potential of AI coding tools may not lie in helping professional programmers write faster, but in enabling professionals who never imagined they could code to start solving their own problems.

The "Citizen Developer" is a concept that emerged from the low-code/no-code movement, referring to business users without formal software engineering backgrounds who build applications using visual platforms or AI-assisted tools. AI coding assistants push this concept further — users don't even need to drag and drop components; they simply describe what they need in natural language and receive runnable code. The deeper implication is that the "informal requirements" long locked inside domain experts' heads — needs too niche, too custom, too small to justify commissioning an engineering team — now have a low-cost path to implementation. The risks are equally worth noting: code written without engineering discipline has inherent weaknesses in maintainability, reproducibility, and security. Under the peer review and data reproducibility standards of medical research, these weaknesses could be significantly amplified.