Signal 013 — April 7, 2026

The Weight of the Machine

9.4 million entities. 44 sources. 189 sitemaps. 20 federally excluded healthcare providers still actively billing Medicare. The machine crossed a threshold today — and the server almost didn't survive it.

The day the database outgrew the hardware, fraud surfaced in the data, and the machine learned to run itself

Wave 7 Started at 9am and Never Asked Permission

3.8 million New York corporations. 165,000 Chicago business licenses. 211,000 NY Medicaid providers. All ingested, normalized, deduplicated, and indexed before noon. The database crossed 8 million entities while I was still on my first coffee.

3.8M
New York corporations
165K
Chicago business licenses
211K
NY Medicaid providers

Then Wave 8

CA Medi-Cal. CMS National — all 50 states, 2.8 million provider records downloaded as a 668MB CSV because the API would have taken 70 hours. Ingested in 23 minutes at 2,400 rows per second.

9.4M
Total entities
13.3M
Total records
44
Sources indexed

189 sitemaps. The machine didn't ask if I was ready.


20 Excluded Providers Still Billing Medicare

Then the fraud signals.

20 NPI-verified healthcare providers federally excluded by HHS OIG — banned from billing Medicare and Medicaid — still enrolled as active providers in national CMS databases.

Not allegations. Not estimates. Exact NPI number matches across two federal databases. The same provider number appearing in the exclusions list and the active enrollment registry simultaneously.

One of them was excluded three weeks ago. Still billing.

I tried to report it to the HHS OIG fraud hotline. The form was so deliberately complicated I gave up. That tells you everything you need to know about who the system is designed to protect.

The fraud signals are ready. The methodology is documented. The report is built. 20 providers. Every NPI verified. Every source a .gov URL. The right journalist gets this and it becomes a story. The story becomes inbound. The inbound becomes the business.

Then the Crash

9.4 million entities. 11GB database. 4GB RAM server. The sitemap generation ran while the API was serving traffic. CPU hit 100% for 50 minutes. The API died. The site returned 502. Real users saw it.

That's the moment you find out if you built something real or something fragile.

23 Minutes to Find the Root Cause

One performance diagnosis by Claude Code found the root cause in 23 minutes — a LIKE fallback scanning 9.4 million rows when the range scan had already found everything.

67s
Search before
37ms
Search after

$48 Per Month for 9.4 Million Entities

4GB RAM to 8GB RAM. $24 more per month. The entire database now sits in memory. Cold starts gone. Every search instant.

$48
Total monthly infrastructure
9.4M
Entities searchable
37ms
Search response time

Nobody is doing this because they think it can't be done cheaply. We proved otherwise.


The Machine Runs on Its Own Schedule Now

It harvests at 4am. It regenerates sitemaps at 6am. It restarts itself if it crashes. It doesn't need me to be watching.

That's the whole point.

Christian Fuhrmann
CEO & Founder, CFAISolutions
Built with AI. Documented in public. Nothing hidden.
cfva.ai · cfaisignal.com · cfaiscore.com