Pipes beat models 🔧, Makary gets fired 🏛️, AI enabled Street Medicine
The Pipes Are The Product
A healthcare AI founder shipping 90,000+ payer calls a month posted the diagnosis: The model doesn’t matter if no one can use it.
The thesis: a 5T parameter model does not move a security review by a single day. Integration is the product — not the phase before it.
Clinicians grade AI by whether it is in workflow rather than the number of parameters. CSV over SFTP this quarter beats Epic write-back next year.
😤 “But the model is what creates the value.” The model creates the capability. SOC 2, the BAA, the service account, and the native write-back queue create the deployment. Capability without deployment is a demo.
💡 80/20: Before your next build sprint, list every non-engineering dependency between you and production: security review, governance committee, service accounts, data agreements. That list is your actual product roadmap.
Makary Out at FDA
Trump signed off on firing FDA Commissioner Marty Makary, a little over a year into his tenure. Clashes over vaping policy, mifepristone, and vaccine messaging all contributed.
For health tech builders, this matters because Makary was the MAHA-aligned commissioner. He pushed the CMS/FDA RAPID pathway for AI medical devices. The replacement will inherit an FDA with 1,430+ authorized AI tools and no clear reimbursement pathway for most of them.
The regulatory weather just got harder to predict. Any builder in the SaMD pipeline should watch who gets named next and whether the RAPID pathway survives the transition.
😤 “FDA leadership changes happen all the time — this won’t affect my product.” If your product touches the 510(k) or De Novo pathway, your reviewer’s priorities just shifted. Changes at the top cascade into review timelines, guidance documents, and enforcement posture. It matters.
ScopeAI Takes AI to the Street
Akido Labs published results from deploying ScopeAI for street medicine — community health workers using an AI-guided intake and diagnostic support tool to serve unhoused patients in the Bay Area.
92% top-three diagnostic accuracy. Patients initiated on MOUD within four hours of first contact. The team increased patient volume without additional staffing costs.
This is a clinician-builder story. ScopeAI supports the full arc of a visit — intake, investigation, structured assessments for licensed provider review. The AI isn’t replacing the clinician. It’s giving CHWs and MAs the clinical scaffolding to work at the top of their scope.
Anthropic Can Now Read Claude’s Mind
Anthropic released Natural Language Autoencoders — a method for decoding Claude’s internal activations into plain English.
The findings are striking: Claude suspects it’s being tested 16-26% of the time but admits it less than 1%. Auditors caught misaligned motivations in 12-15% of cases.
For clinical AI, this is the interpretability tool might be what we’ve been waiting for.
Cloudflare Cuts 1,100 Jobs for “AI-First” Restructuring
Cloudflare eliminated 1,100 positions while citing a shift to an “agentic AI-first operating model.” Revenue per employee is up ~600% over three years.
🛠️ From the Workbench
Stedi: Build an Insurance Verification App in 30 Minutes
Stedi published a tutorial on building an insurance verification app using Claude Code and their eligibility check API.
Real-time eligibility checks against 2,100+ payers. Free Basic plan, MCP server for agent integration, structured HIPAA-compliant response parsing.
⚠️ Verify: Stedi’s API handles the payer connection, but if you’re feeding real patient data through Claude Code (or any third-party agent), you need a BAA in place. The tutorial is great for learning on test data. Production use requires a lot more.
💡 80/20: Sign up for Stedi’s free plan and run a test eligibility check with their sandbox payer. The experience of seeing structured 271-response data come back in seconds — instead of navigating a payer portal — is the demo you need to understand why RCM automation is a real product category.
🧰 Builder’s Tip
Map Your §170.315(b)(11) DSI Source Attributes Before Your Next CMIO Meeting
If you’re building any clinical AI tool that touches a certified EHR, the Decision Support Interventions (DSI) Source Attributes rule has been live since January 2025.
The rule requires certified EHR modules to display source attributes for any CDS or AI recommendation: who built it, what data it was trained on, how it was validated, known limitations, and update frequency. Five sections, 29 attributes total.
Here’s the workflow pattern: open the ONC §170.315(b)(11) criteria, pull the 29 attribute names into a spreadsheet, and fill them in for your product. Two weeks of part-time work. The output is a one-pager that answers every question people will ask about your AI’s provenance, and it’s the procurement-readiness artifact that separates serious builders from demo-stage projects.
What are you building this week? Reply and tell me — I read every one.
— Kevin