I’m a licensed Medicare supplement and Part D agent. 8 states. Solo practice. This is the side of my business that actually pays the bills right now.
On April 1, 2026, I wrote the first line of code for the Medicare Agent. 90 commits later, I had a working system. Not because I wanted to build software. Because the industry kept taking away the tools I needed to serve clients.
The problem
Carriers pull their agent tools constantly. One quarter you have a plan comparison engine. Next quarter it’s gone or locked behind a new contract tier. Commissions shrink. The platforms you relied on get acquired and shut down.
I got tired of depending on tools I didn’t control.
What I built
A Next.js application with:
- RAG over Medicare knowledge base. A conversational research assistant that searches official CMS data, not carrier marketing materials. The bias is intentional. CMS is the source of truth.
- Client CRM. Basic but mine. I know where every client is in their enrollment journey.
- Plan comparison engine. Side-by-side plan analysis using actual cost data.
- Drug cost modeling. What will this formulary actually cost the client over a year.
- Enrollment package generation. PDFs I can send directly.
- Appeal letter generator. When a claim gets denied, the system drafts the appeal referencing specific CMS guidelines.
- CMS annual update tracking. When premiums, deductibles, and IRMAA brackets change each fall, the system flags what needs updating.
All built with Claude Code. SQLite for data. Qdrant for vector search. Python for PDF generation.
The practitioner angle
This isn’t a tech company building for the insurance industry. This is an insurance agent building his own tools because the existing ones weren’t good enough.
That distinction matters. I know what data I actually need to serve a client. I know which CMS tables matter. I know the MRI story. A client got an MRI denied, forced to do a CT scan first, waited 6 months, paid extra, and worried the whole time. With Original Medicare and Plan G, that MRI happens at the next appointment.
The tool encodes that kind of practical knowledge. Not theoretical. From doing the work.
90 commits in 28 days
One person. One repo. One terminal running Claude Code.
The velocity is the part people don’t believe. But it’s real. AI coding tools don’t make you a better engineer. They make a practitioner dangerous enough to build the thing they actually need.