Helium Scribe product mockup
challenge
Imagine visiting a doctor in a busy Nigerian hospital. The consultation ends. You walk out. The doctor doesn’t move on to the next patient. They sit down and spend the next 20 to 40 minutes typing up everything that just happened from memory.
Now imagine that doctor has twelve patients today.
That’s not an edge case. That’s the daily reality for most Nigerian clinicians. And the problem isn’t that doctors are slow typists. It’s that the workflow forces them to be fully present with a patient and then immediately reconstruct everything that happened, in writing, from memory.
By the time they sit down to type, details are already fading. Notes become generic. Clinical nuance gets flattened. And “I’ll come back to finish this later” very often means it never gets finished.
impact at a glance
5,000+ hours saved
Documentation time recovered in a single hospital
15 physicians
In one hospital alone
5 hospitals tested
Across Nigeria before launch
We talked to 5 doctors across 5 different hospitals in Nigeria
Before building anything, we spent time inside hospitals. Five doctors, five different hospitals, real clinical environments.
Here’s what we found.
Incomplete notes are not just an admin problem. They’re a clinical risk. A patient returns two weeks later, a different doctor covers the shift, and the record doesn’t tell the full story.
Existing EMR tools didn’t help. They were built for recording decisions that had already been made. Nothing existed for the moment the decision was still happening. Every doctor we spoke to described documentation as the part of their job that followed them home.
The core problems
Documentation happens after the patient leaves — Doctors reconstruct conversations from memory minutes or hours later, by which point the clinical detail has already degraded.
Notes are incomplete or generic — Under time pressure, doctors default to shorthand. The nuance that matters for follow-up care doesn’t make it into the record.
It compounds across the day — One incomplete note is a small problem. Twelve incomplete notes across a full clinic day is a clinical risk that accumulates quietly.
Why AI and not just better templates?
We considered structured input fields. Helium Health already had them. Doctors abbreviated aggressively and still spent the same amount of time after consultations. The problem wasn’t the form. It was the act of authoring.
Voice-to-text alone wasn’t the answer either. A raw transcript of a clinical consultation is noise. Patient tangents, small talk, reassurances mixed in with actual clinical content. Handing a doctor a raw transcript just moves the problem instead of removing it.
Here’s what AI actually unlocked.
The ability to listen to an unstructured clinical conversation, understand what mattered clinically, and produce a structured note without the doctor doing any of that manual work. That requires contextual language understanding. Medical domain knowledge. The ability to tell the difference between a symptom and a pleasantry. Rules-based systems cannot do that. Keyword matching cannot do that.
Here’s how we designed it
The flow was straightforward in concept. Doctor starts the consultation, Helium Scribe listens, transcribes in real time, then generates a structured clinical note the moment the consultation ends. By the time the patient walks out the door, a draft note is already waiting.
Notes were generated in SOAP format. Subjective, Objective, Assessment, Plan. That wasn’t arbitrary. SOAP is how Nigerian clinicians are trained to document. Generating output in that format meant doctors were reviewing something that matched their mental model of what a note should look like.
Where the AI was uncertain, unclear audio, ambiguous language, insufficient context, those sections were visually flagged. The doctor’s eye went straight to what needed attention. We surfaced uncertainty explicitly rather than presenting uniform confidence that hid errors.
Doctors could edit any section inline. If a specific section was wrong they could flag it and trigger regeneration of just that section, without losing the parts that were correct. Nothing was committed to the patient record until the doctor explicitly approved it.
Review interface: SOAP sections, uncertainty flags and approval
Then we hit a wall
We built Helium Scribe as a widget on the desktop and web version of the Helium Health EMR. Made sense on paper. Doctors were already using the EMR on laptops in their offices.
But during early testing, the first doctor we visited broke that assumption immediately.
He was an obstetrician. He doesn’t sit at a desk. He moves from ward to ward, patient to patient, with his mobile device. No laptop in sight. And beyond his situation, other doctors started reporting that their laptop microphones were picking up too much background noise from busy clinic environments, making transcription quality inconsistent.
The desktop-only approach wasn’t going to work.
Doctor moving between wards with a mobile device
So we built a mobile companion
The solution was a dedicated mobile app for audio capture, connected to the desktop EMR session via QR code and PIN authorization.
Here’s how it works now.
The doctor opens the patient record on their desktop EMR and a QR code appears. They scan it with the mobile app. Authorization confirmed. The mobile device becomes the microphone, capturing audio cleanly regardless of where the doctor is in the building. The transcription and note generation still happen on the desktop. The mobile device handles the one thing it’s genuinely better at, capturing clean audio in a moving clinical environment.
This meant the doctor could conduct a consultation at the bedside with their phone in their coat pocket and return to their desk to find a completed draft note waiting.
Desktop EMR → QR code pairing → Mobile audio capture → Desktop note generation → Doctor review
The consent problem nobody talks about
The most painful thing we had to solve wasn’t the AI. It was consent.
The patient is being recorded during their consultation. They never touch the app. There’s no screen for them to tap. But their voice, their symptoms, their medical history are all being captured.
We needed verbal consent collected by the doctor at the start of every consultation. The design challenge was giving the doctor a clear prompt to do this without it becoming an awkward compliance step that broke the natural flow of the consultation.
We went through several versions before we landed on something that worked. It sounds like a small problem. It genuinely wasn’t.
One thing I made sure
Every consultation that went through Helium Scribe was tagged in the patient record as AI-assisted. Audit teams could review AI-generated documentation separately. And it made the doctor’s approval feel like a real act of responsibility, not a rubber stamp.
Results and impact
5,000+ hours of documentation time recovered for 15 physicians in a single hospital.
Amplitude showed adoption climbing steadily after launch and continuing to climb. That kind of curve means the product changed a habit, not just introduced a novelty.
The feedback from doctors was consistent. Time saved, yes. But more than that, they described feeling less drained at the end of a clinic day. The documentation work had been reduced to a review exercise.
That shift, from authoring to reviewing, was what the product was actually for.
What I learned
The hardest part of designing Helium Scribe wasn’t technical. It was clinical.
Understanding the difference between a note that was technically complete and one that was clinically useful required real time inside hospitals. Early AI outputs were grammatically correct and structurally valid but experienced doctors described them as hollow. Like they’d been written by someone who had read about medicine rather than practiced it.
What surprised me most was how much the trust design mattered relative to the accuracy design. Doctors were willing to tolerate a draft that needed some editing. They were not willing to tolerate a system that hid its uncertainty or made it hard to override. The confidence flagging feature almost got cut because it added visual complexity. It turned out to be one of the most cited reasons doctors trusted the product.
The mobile companion is what I’m most proud of. Not because it was technically impressive but because it came directly from watching a real doctor move through a real hospital. The obstetrician who broke our desktop assumption in the first testing session gave us a better product than the one we had planned to build.
If we were doing this again I would treat Nigerian clinical language as a first-class design constraint from day one. The way doctors actually speak, mixing English, Yoruba, Igbo, Pidgin, local medication names, informed every part of the AI design. It should have been the starting point, not something we accounted for mid-build.
