← Work
MDaaS Global · BeaconHealth

Hazel AI

Real-time consultation AI. From patient complaint to treatment plan in one session.

Role

Tech Lead & Senior Product Designer

Timeline

2024 — Present

Team

Solo designer, engineering, clinical team

Platform

Web app (clinician-facing)

Cover image — Hazel AI consultation interface

The Problem

Doctors were drowning in documentation.

Primary care consultations are cognitively demanding. A doctor must listen to a patient, ask follow-up questions, recall relevant medical history, cross-reference diagnostics, form a diagnosis, and document everything, all at once.

The documentation burden alone consumes a significant portion of each consultation. Doctors type notes while patients are still talking, which fragments attention and reduces the quality of the interaction. After the appointment, they often spend additional time completing records they could not finish in the room.

The goal was to give doctors the mental space to focus entirely on the patient. Let the system handle the structured output.

The cognitive load of a primary care consultation

The Approach

Listen first. Structure later.

Hazel AI listens during the consultation and transcribes in real time. The interface had to feel nearly invisible while the session was active, present enough to confirm it was working, unobtrusive enough not to distract the doctor or unsettle the patient.

The system cross-references the patient's medical history and any diagnostic results already in Olewerk as it listens. When the consultation ends, it generates a structured clinical summary and presents it to the doctor for review before producing a treatment plan.

The three states: listening, generating, reviewing

The handoff moment was the most critical design decision. This is where the doctor reasserts authority and the AI's output becomes clinical record. The approve action had to feel deliberate, not accidental. The summary had to be scannable, not overwhelming. Every field had to be editable so the doctor could correct the AI without friction.

Designing that handoff clearly and confidently was what determined whether pilot users trusted the tool or abandoned it.

Summary review screen

Treatment plan approval state

Outcomes

Less time writing. More time with patients.

Live

With pilot users across BeaconHealth primary care

Significant

Reduction in post-consultation documentation time

More time

Spent on patient interaction per session

Hours saved

Per doctor per week on record-keeping

Hazel AI in use during a primary care consultation

Reflection

What this project taught me.

The hardest part of designing ambient AI tools is making the interface trustworthy during the moments it is invisible. Doctors needed to know Hazel was listening accurately without that awareness becoming a distraction.

Designing the listening state, subtle enough not to intrude, clear enough to confirm it was working, was a problem I had not encountered before. It changed how I think about AI presence in UX: sometimes the best interface is the one you barely notice is there.

Next Project

Motherboard Platform Redesign

Read case study →