AI assisted clinical session notes, built to be trusted
Chose clinician verified drafts over autonomous notes, and pt-PT clinical fidelity over a faster launch on a generic model.
Star Mountain is a clinic management SaaS I am building solo for psychology and psychiatry practices in Portugal. It covers scheduling, integrated teleconsultation, electronic clinical records, AT compliant billing, and RGPD compliance. The feature explored here is optional AI assisted session notes in European Portuguese.
Clinicians lose meaningful time writing notes after every session. It is administrative drag that eats into clinical hours and feeds burnout. A language model can draft those notes, but this is the opposite of a low stakes use case: the data is sensitive mental health information under RGPD, the audience is clinicians who will not trust output they cannot verify, and a fabricated clinical detail is not a cosmetic bug, it is a safety and liability problem.
- Draft, never decide. The feature is framed as a first draft accelerator, not a source of truth. Nothing enters the clinical record until a clinician reviews, edits, and signs off. That sign off is not a nicety, it is the core safety mechanism of the product.
- Human in the loop by design. Generate, then clinician edits, then explicit approval, then the record. I placed the human at the one point where a mistake is caught before it becomes permanent, rather than treating review as an optional afterthought.
- A working definition of good enough. Quality was not defined as elegant prose. It was defined as a faithful, editable starting point that saves time without inventing clinical facts. I evaluated against clinician edit effort and time saved, with a hard rule that no invented findings are acceptable, which matters more than fluency.
- European Portuguese, not a translation. Clinical register in pt-PT is distinct from Brazilian Portuguese and from translated English. A note that reads foreign erodes trust instantly, so I chose to invest in pt-PT specific prompting and evaluation rather than ship faster on a generic multilingual default.
- Privacy as a purchase decision. For a clinic, how sensitive health data is processed is a buying factor, not just a compliance checkbox. I designed around data minimisation, clear consent, and explicit processing boundaries, and made those legible to the clinic rather than buried.
- Latency budget spent on quality. Notes are generated after the session, not live, so the latency budget is generous. I used that headroom to prioritise accuracy and cost control instead of chasing real time speed the use case does not need.
The product thesis is validated where it counts: the differentiator is not that notes are generated, it is that they are generated in a way clinicians will actually adopt, in their own clinical language, without asking them to trust a black box. The trust model, human sign off plus pt-PT fidelity plus visible privacy, is the moat, and it is a harder thing to copy than the generation itself.
An early working prototype of the clinic dashboard. The day timeline re-divides into 15 to 60 minute slots, with a live clock, appointment counter, status KPIs, per professional load, a waiting list, and a projected revenue snapshot. Patient names are abbreviated by default, so privacy is the resting state rather than a setting. The interface is in European Portuguese by default, with a PT/EN toggle for English readers. Try the interval control, the professional selector, and the language switch.
The lesson that carries into any AI role: with a probabilistic product, you design the trust model before you optimise the output. Deciding where the human sits and what the model may touch is the senior work. Tuning the prompt comes after.