15 Nov
|
Curaeon Health
|
The Hills Shire
15 Nov
Curaeon Health
The Hills Shire
Curaeon is practice management software for Australian general practice, built in Sydney — with an on-premises AI scribe that drafts the clinical note without any consult audio, transcript or clinical content ever leaving the practice. Data sovereignty isn't a feature of the product; it is the product. You'd join before the first paying clinic — the interesting part, and the honest risk.
You own the appliance: the ASR and the drafting model that turn a consultation into a clinical note, running on hardware in a doctor's office.
Why this role exists
Everything a scribe normally does easily, we do the hard way on purpose. We can't call a frontier API. We can't ship audio to a datacentre for batch processing. We have one GPU box per practice, a doctor waiting at the end of a consult, and a note that has to be good enough for a clinician to sign their name to.
Today the pipeline is faster-whisper for transcription and a local LLM for drafting, wired through a Python worker that pulls audio from an on-LAN object store and posts partial transcripts back over SSE. It works. It is nowhere near as valuable as it needs to be, and there is no one whose job it is to make it better. That is this job.
What you would work on
Note quality, measured rather than felt. Eval sets that tell us whether a change improved anything — across accents, specialties, interruptions, three-way consults with a family member in the room, and the Australian clinical vocabulary that general‑purpose ASR mangles.
Latency and footprint. The draft should be waiting when the doctor stops talking. Quantisation, batching,
model selection, KV cache behaviour, and what actually fits in the RAM of a box a practice will pay for.
The safety boundary, as an engineering problem. The model must not produce a diagnosis. We enforce that with a prompt, a schema, a deny‑list validator and a release gate that fails the build. All four are yours to strengthen — the interesting failures are subtle, like a model that writes “consistent with” instead of naming the condition.
Speaker handling. Knowing who said what changes a note from a wall of text into a record.
Choosing what runs. Local model landscape, licence terms, what we can fine‑tune, what we should not.
What we need
Real experience getting models to run in production under hardware constraints — quantisation, serving stacks (vLLM, llama.cpp, TensorRT‑LLM or similar), and the profiling to know where the time went.
Python, and enough software engineering to own a service rather than a notebook.
Evaluation as a discipline. If your instinct on “is the new prompt better” is to build a dataset rather than read ten outputs, you're who we want.
ASR experience, or the appetite to own it quickly.
Nice to have
Clinical NLP, or any work where being wrong had consequences.
Fine‑tuning small models — LoRA and friends — with a clear view of when it's worth it and when a better prompt was the answer.
This is not the job if
You want to train foundation models, or to build on the newest hosted API. Neither happens here, and the constraint is permanent — it's what we sell.
📌 ML/LLM Engineer (The Hills Shire)
🏢 Curaeon Health
📍 The Hills Shire