The idea
I generally have interest in applying new technology to my professional interests, and often try out tech "hands-on" to help me understand it better (and potentially to help the patients and families I care for).
Ongoing and past "tech" projects:
- Clinical advisor to Fiber Health, maker of Pooficient (ongoing, financial interest)
- Manuscript in progress about use of GAI in research (ongoing — exciting!)
- Publications about technology and GAI relevant to my profession/Pediatric Gastroenterology (see Google Scholar or PMIDs 40752924, 39243338, 37079871, 36976575, 37168904, 34105388) and a chapter in the upcoming "Authoritative Guide to Chronic Pediatric Abdominal Pain" (Saps, Benninga, eds., Academic Press)
- Created an NFT ("work uniform for a pediatric gastroenterologist") for fun, and it is rightfully valued at $0
Despite interest and some knowledge about digital technologies, I have no coding experience since working on BASIC in the Apple IIe lab in 1989. Vague memories of moving a triangle around the screen and figuring out how to make the background flash rapidly across the color spectrum.
Recently I started to use Claude (Sonnet 5, High) and Claude Code to address a project that was floating around for the last year.
- OpenPayments is a CMS database with reported industry→clinician payments for research and other activities (consulting, etc).
- The OpenPayments reporting system is not perfect, and this project was not intended to fix it, nor intended to comment on individual activities — only as a way to visualize data more easily.
- The question I considered was whether Claude Code could help me visualize payments to professional organization leadership in aggregate (all the officers at once).
- The problem with doing this manually, as far as I tried, was that it required individual identification of leaders and input of their information into the (pretty well-made) online OpenPayments portal.
- OR the entire database can be downloaded and a local search run, but it is unwieldy/huge and still requires re-download when new data is published.
- So, could I code a process to obtain this data and visualize it, and how could Claude Code assist?
I wanted to see, in one place, how much industry money flows to the people leading national pediatric subspecialty societies — NASPGHAN, ASPHO, PIDS, and others — broken out by general versus research payments, using CMS's public Open Payments (Sunshine Act) data.
I decided to see how far I could get by describing what I wanted rather than writing the code myself — working with Claude first in a chat interface, then handing off to Claude Code once the project needed real internet access.
Phase 1: Prototyping with fake data first
Before touching any real data, I built a working dashboard using realistic placeholder numbers: a year-comparison bar chart across organizations, and a per-organization trend view over time. That let me react to the actual look and feel of the thing — the layout, the colors, the chart types — without worrying about the actual data.
Phase 2: Defining "who counts"
This turned out to be harder than the charts. "Officers of pediatric subspecialty societies" needed to be parsed — who of the listed officers should be included (at-large, past-president, etc). I settled on full boards/councils, tracked across their complete payment history in the time assessed, regardless of exact tenure with the organization. This is imperfect, but a reasonable tradeoff given that organizations only publish current leadership online.
Phase 3: Collecting rosters
I gave Claude the leadership pages for seven organizations. It pulled 65 individual names and titles, and there was one major error related to website layout — the names pulled were completely incorrect, but Claude was able to fix the problem when redirected. Another example of using AI as a tool, not as the end-all (humans are still important for the process).
Phase 4: Matching names to a stable identifier
Physician names alone aren't enough to pull clean records, especially when the name listed on an org webpage might not equal an official or legal name. CMS explicitly recommends matching by NPI (National Provider Identifier) instead. We piloted this on the smallest board first (7 people) before scaling up — matches came back with old addresses, with multiple potential matching NPIs (same name), with "Pediatrics" but not the pediatric subspecialty listed in the taxonomy identifier. This is an issue with scale and reproducibility, and requires a bit of manual intervention. I told Claude to exclude officers specifically if they were (1) not practicing in the United States (no NPI, no Open Payments record), or (2) not an MD/DO or APRN (several PhDs, again without an NPI or Open Payments record), and to use taxonomy matches against professional organization to deduce the right person when multiple name matches existed. Even though Claude worked through all of this, it was definitely helped by a human who understood some of the rules to apply.
Phase 5: Hitting a real technical wall
Trying to pull the actual payment records into Claude hit a limit — the Claude interface stopped the type of approach we were taking (the way it was conducting API queries). Rather than work around it, we moved everything — rosters, matched NPIs, all the problems found so far — into a brief for Claude Code. Claude Code has "real" internet access and could appropriately contact and use the Open Payments API and the NPPES (NPI) registry as needed. Plus a lot of debugging that I don't really understand. In this case, Claude saw its own limit and directed me to Claude Code appropriately. This also resulted in installing some software to get the job done (Claude Code, Python, Git).
Phase 6: Finishing in Claude Code
This was in a language I have minimal proficiency in. Summarized in a project brief that Claude Code created — but actually a very nice step-by-step process. I asked questions about terms, steps, where to find files — probably very basic for someone who codes, but Claude Code was patient with me, and I know a bit more now than when I started.
Phase 7: Going live
The last step was turning the working dashboard into something publicly reachable. Since it's free and seemingly very functional, I used a GitHub Actions workflow that automatically builds and deploys the site to GitHub Pages every time I push a change. Claude told me how to do this step by step, and it only took a couple of tries. After the second failure I asked Claude Code where I went wrong, and it found the error (something about a .git repo and a deploy.yml it didn't recall setting up, and a generic Node.js/Vite template that didn't match the project at all), then fixed it.
What I'd tell someone starting out
Go for it. The best part (for me) about GAI tools and other digital technology is that they are breaking down the barriers. Information is more accessible than ever, and while technical skills are still invaluable(!!), modern tools make it possible for "tech-interested" people like myself to do work that would have been long and difficult just a few years ago. From start to finish, this took two weekdays of spare time, only. If you have an idea, ask GAI how to accomplish it. Then ask more questions.
Keep going down the rabbit hole and you might be surprised what comes out the other side (my cliche, not Claude's).