JournalChangelog
Scribara/Journal

Notes from the team building it.

Clinical AI, agent reliability, healthcare workflow, the building. Long-form, infrequent, and written by people in the work, not the marketing team.

May 24 · Renata Liang, MD · Clinical

Why "ambient" is the only architecture that survives the exam room.

Push-to-talk dictation lost the room a decade ago. Form-driven scribe UIs lose it every visit. The only model the patient doesn't notice is the only model the clinician will tolerate.

May 17 · Jonas Krieger · Engineering

Verifier models, explained — and why ours is a separate net.

Self-consistency is not enough at clinical reliability levels. We run a second, independently-trained model whose only job is to disagree. Here's the architecture.

May 9 · Adit Mahajan · Strategy

Specialty is the wedge. Primary care is the trap.

Generalist documentation is solved-enough. Specialty documentation is hard, valuable, and budget-funded. A primer on category construction.

May 1 · Sara D'Avella · Design

Designing for the room you're not in.

Three principles that shaped the in-visit UI — and one mistake we shipped twice before we caught it.

April 22 · Engineering team · Research

How we evaluate model upgrades on PHI without exposing PHI.

A walk through our production-shadow eval pipeline, including the cryptographic and contractual primitives that let us measure quality without ever decrypting a patient name.

April 14 · Renata Liang, MD · Clinical

The CPT modifier that's worth $40,000 to your practice.

Modifier 25 is the most under-applied code in specialty medicine. Here's the documentation pattern that supports it, and how Scribara catches it 96% of the time.

April 7 · Jonas Krieger · Engineering

RAG over payer policy: a hard retrieval problem.

Payer prior-auth policies aren't documents — they're decision trees with citations. We built a hybrid retriever that respects the tree.

March 31 · Adit Mahajan · Strategy

What "AI replaces labor, not seats" actually means in pricing.

Three pricing experiments. One that worked. Why the seat-based AI startup is a category mistake.

March 24 · Security team · Trust

ISO/IEC 42001, and why we sought it before SOC 2 Type II.

The new standard for AI management systems matters more than the certification industrial complex realizes. Here's why we went first.

Subscribe

The Tuesday note.

One short essay every Tuesday morning. From the team building Scribara. No newsletter, no ads, no growth-hack hooks.