Vlad Gavrilov

Adaptive Learning: Encoding Clinician Edits as Memory

Clinicians edit their AI-generated notes for many reasons: adding or removing content, fixing spelling, and preferences for structure, wording, or terminology. Some of these edits are contextual, applying in some situations but not others. The data is noisy, but inside it is useful signal that can be used to predict and make these edits before a clinician ever sees the note, cutting the time spent editing and giving it back to the patient.

To do this we built a custom agentic workflow that currently processes hundreds of thousands of notes and edits in production, and encodes them into reusable memories per user that correct their future notes. Each memory carries a sense of when it should and shouldn't apply, which is what keeps edits precise. In this talk I'll walk through the harness design (tools, memory, evals, self-healing), as well as some key lessons from the failure modes and experimental results that helped shape the build.