Halvia · AI systems for healthcare
I'm Shaun Finch, a UI/UX and motion designer in Manchester working on AI in health and science. I design, build and set up AI tools around how doctors, nurses, coders and patients actually work, with a person always making the final call. Seven of them are working software you can try below.
Synthetic patients onlyNothing leaves your browserA clinician approves every output
Experience across health, broadcast and sport
Case study · CHKS
A product explainer for CHKS's AI clinical coding system, used across the NHS. The film had to make a technical product feel simple and trustworthy for coding teams: what it does, how it fits their day, and why the coder stays in control.
Coding teams are experts. The film needed to show what the AI does without overselling it, and leave viewers clear that coders review every suggestion.
Turning the product story into animated sequences: the coding journey, the interface moments and the closing message.
One film covering coding status from not coded to fully coded, local rules, readiness for ICD-11 and the experience behind the product.
Patient explainer · AI video
A short vertical film for patients to watch after a GP or minor injuries visit. It shows what's happening under the skin, the cold compress, pain relief, the warm compress after a couple of days, and the warning signs that mean coming back. Clinicians explain all of this in the room, but patients often forget it by the time they get home.
AI-generated video. A concept to show how generative video could support patient information. In real use, every script is checked by clinicians against current guidance before a patient sees it.
What I do
Most AI projects in health don't fail on the model. They fail because the tool doesn't fit a twelve-hour shift, nobody trusts the output, or patients can't read what it says. That gap is where I work.
UI/UX for AI systems built around doctor and nurse ease of use: clear screens, explainable flags, and designs tested with the people who'll use them.
Front-ends and tools that actually run, like Halvia Ward, so a ward or team can try the idea before anyone signs a contract.
Connecting AI into existing workflows and checking it keeps working: evaluation of model outputs, audit trails and handover documentation.
Seven tools · working software
Each tool is a working concept for one job on a ward, in a records office or in clinic. They share one synthetic ward, so recording observations in one updates the ward board in another. Use them here, or open the full-screen workspace.
Tools
halvia://ward/overview
Every bed's NEWS2 score, trend and next observations on one board. Press Live and the demo clock runs: observations arrive as often as NEWS2 says they should, and a patient who's getting worse rises to the top with an escalation ready to review.
Nurses in charge piece together who's sickest from paper charts, e-obs screens and handover notes, often while doing their own patients' care.
A live board fed by the e-obs system, sorted by risk, showing overdue observations and trends, with one tap into the chart and an escalation draft.
Overdue observations per shift, time from a trigger score to review, and how often staff use the board at the start of a shift.
halvia://ward/early-warning
NEWS2 is the early warning score used across UK hospitals. Scoring it is simple; noticing the trend across a busy shift, and escalating clearly, is the hard part. This scores observations live, charts the trend and drafts the escalation for the nurse to check and send.
Deterioration is often visible in the observations hours before a crisis, but it's spread across charts and shift changes. Escalation calls can be rushed or unclear.
A bedside view that scores as obs are entered, highlights rising trends across shifts, and drafts an SBAR escalation the nurse edits and sends from the same screen.
Time from a trigger score to review, missed escalations, nurse-rated clarity of calls, and alert fatigue.
halvia://ward/handover
Paste free-text notes and the assistant sorts them into SBAR (Situation, Background, Assessment, Recommendation), pulls out the to-do list and raises safety flags, including a medicine and allergy clash hidden in Bed 3's notes.
Handovers vary by person and by how tired they are. An allergy query, a near-miss fall or a pending blood test can sit in a paragraph and get missed.
A handover tool that structures notes into SBAR, cross-checks medicines against allergies, and asks the outgoing nurse to approve each section, with a record of what was handed over.
Handover time, items missed in audit, incidents linked to handover, and whether incoming staff feel they have what they need.
halvia://ward/discharge-letter
Discharge summaries are written at the end of a long day and GPs often get them late or incomplete. This turns ward notes into a structured letter: diagnosis, what happened, every medicine change, actions for the GP and what the patient was told, with each line linked back to its source.
Medicine changes and GP actions get lost between ward and practice, which leads to repeat tests, missed follow-up and avoidable readmissions.
A drafting tool inside the discharge workflow that pulls from the notes and drug chart, follows a standard structure, and won't send until medicines are checked.
Time to write a summary, letters sent within 24 hours, medicine discrepancies found by GPs and pharmacists, and GP satisfaction.
halvia://records/clinical-coding
Clinical coding turns each hospital stay into standard codes that drive funding, research and planning. In England that means ICD-10 for diagnoses and OPCS-4 for procedures. This suggests ICD-10 codes, shows the sentence behind each one, and spots conditions that were ruled out so they aren't coded by mistake.
Coders read long, unstructured notes under time pressure. Missed or wrong codes affect funding and the data used to plan, and backlogs build quickly.
A suggestion panel inside the coder's existing system: evidence for every code, negation and uncertainty spotted, and one-click accept or reject that feeds an audit trail.
Episodes coded per hour, audit accuracy, depth of coding, and how often coders override suggestions, and why.
halvia://outpatients/missed-appointments
Missed appointments waste clinic time and delay care. This estimates who might struggle to attend, explains why in plain terms, and suggests help: a reminder, a video slot or a nearer date. It's never used to overbook or move anyone down a list.
Every empty slot is time a patient on the waiting list could have used. Blanket reminders help, but the people most likely to miss often need a different kind of help.
A daily list for booking teams showing who may need support and why, with reminder, video and rebooking actions built in, and freed slots offered back to the waiting list.
Missed-appointment rate by clinic and by patient group (to check it's fair), slots refilled and staff time per list.
halvia://patients/plain-english
Clinic letters are usually written for other clinicians. This rewrites one for the patient: jargon explained, medicines in a table and next steps up front. It measures the reading age before and after, and can read the letter aloud.
Many adults struggle with health information written at a high reading level. Patients leave clinic unsure what was found or what to do next, and call the hospital to ask.
A letter assistant that drafts a patient version alongside the GP letter, with a glossary, medicines table, reading-age check, translation and audio.
Reading age of letters sent, patient understanding, follow-up calls to the clinic, and the extra clinician time per letter.
Halvia Ward, full screenAll seven tools in one workspace with an activity log, keyboard shortcuts, print and copy. Six synthetic patients, ready to use.
Open Halvia Ward ↗How the demos work
These tools run on simple, visible rules so you can see exactly why each output appears. In a real project the rules are swapped for validated models and your systems, but the shape stays the same: every suggestion is explained, and a person approves it.
| Tool | In this demo | With real models and systems | Safety gate |
|---|---|---|---|
| Ward overview and early warning | The NEWS2 chart from the Royal College of Physicians, plus a simple trend across recent scores | Fed live from the e-obs system; deterioration models that also use blood results, validated on local data | Thresholds agreed with outreach; a nurse confirms every escalation |
| Shift handover | Keyword rules sort sentences into SBAR and match a list of safety flags | A language model summarises from the record, cross-checked against prescribing and allergy data | Every section approved; each flag shows its source sentence |
| Discharge letters | A parser reads labelled ward notes and builds the letter structure | Drafted from the full record and drug chart, following the national discharge summary headings | Won't send until medicines are checked; the clinician signs off |
| Clinical coding | A dictionary of about twenty ICD-10 codes, with rules for negation | Clinical language processing mapped to the full ICD-10 and OPCS-4, with national coding standards | An accredited coder accepts every code; audit sampling |
| Missed appointments | Hand-set weights in a simple risk formula | A model trained on local attendance data, tested for fairness across patient groups | Used only to offer help; never to overbook or deprioritise |
| Plain-English letters | A glossary and a reading-age formula | A language model rewrites to a target reading age, with translation and audio | The clinician signs off the patient version |
Medtech & biotech software
Hearing aids, prosthetic hands and lab instruments all run on software, and so do the apps, dashboards and set-ups around them. These are real roles in medtech and biotech teams. I work on the design and software side, alongside your engineers, clinicians and regulatory team.
Users change volume, programs and streaming from a companion app. Audiologists use fitting software to set each device to the person's hearing test.
Myoelectric prosthetic hands are set up through apps that choose grip patterns and tune how sensitive they are to muscle signals (EMG). Prosthetists use clinic software to calibrate them.
When software is itself a medical device, it's built under a defined lifecycle and a usability engineering process, with every design decision documented.
An independent look at an existing codebase before an audit, a funding round or a handover to a new team.
Labs run on instrument exports, analysis pipelines and a lot of spreadsheets. Good tools make results quicker to read and easier to trust.
Getting systems talking to each other, then keeping them working once real staff and patients are using them.
The standards listed are the ones this work sits within, so it fits how your quality team already operates. Certification and regulatory sign-off stay with your quality and regulatory team.
Where this comes from
I've spent my career making complex things clear on screen, and now focus on clinical and scientific products.
Explaining a clinical coding product used across the NHS to the people who use it, plus interactive UI built with Framer and Claude Code. Watch it ↑
Clearer scientific data visualisations, and medical and scientific explainer video.
Motion and digital design for broadcast and entertainment, including EastEnders and iPlayer at the BBC.
Decks & research documents
Clinical and research meetings move fast. I design decks, case-study PDFs and posters that make the finding clear in the first minute, for grand rounds, MDT and governance meetings, journal clubs, funding bids and conferences.
Samples use illustrative data. Real projects are built from your data and checked by your research team before they're presented.
How a project runs
Start with one job and one team, prove it helps, then grow it. Each step ends with something you can see and decide on.
Shadow the team, map the workflow and agree the one problem worth solving first.
1 to 2 weeks · workflow map, success measuresWorking software like Halvia Ward, tested with staff on synthetic data.
2 to 4 weeks · prototype, feedbackA small, supervised pilot with your clinical safety officer and IG team, measured against the agreed outcomes.
Agreed with you · safety case input, evaluationHand over designs, code, training and an evaluation plan, so the tool keeps working after launch.
Ongoing · docs, training, monitoringPrinciples
Six rules every Halvia project follows, written for clinical safety officers and IG teams as much as for designers.
AI drafts, flags and suggests. A named member of staff approves anything that affects care, and the interface makes that step obvious.
Every flag links to the sentence, number or rule behind it. If staff can't see why, they can't trust it or challenge it.
Least data, for the shortest time, in the right place. Designed with UK GDPR, data protection impact assessments and the NHS Data Security and Protection Toolkit in mind.
Hazards logged from the first sketch, so your clinical safety officer can work to DCB0129 and DCB0160, and the DTAC assessment has what it needs.
Plain English, reading-age checks, WCAG 2.2 AA contrast and keyboard support, and layouts that work on a ward PC and a phone.
Model outputs are scored against clinician-written rubrics before and after launch, and checked for fairness across patient groups.
AI agents & automation
Clinicians lose hours a week to inboxes, letters, rotas and reports. I set up AI agents and automations that take on that admin, connected to the systems you already use, with a person approving anything that reaches a patient or a record.
Sorts shared inboxes, drafts routine replies and flags urgent referrals for a clinician to see first.
Drafts clinic letters, discharge summaries and meeting minutes from notes, ready for sign-off.
Prepares clinic lists, chases missing results and suggests cover for gaps in the rota.
Searches and summarises papers and guidelines with their sources, for journal clubs and protocol reviews.
Pulls the numbers into weekly dashboards and audit reports, and chases the data that's missing.
Approved reminders, preparation instructions and plain-English follow-ups, sent at the right time.
Built with tools includingClaudeClaude CoworkGrokyour existing systems
Patient data stays protected. Patient-identifiable information only goes into tools your information governance team has approved, under a data protection impact assessment. Agents work with the least access they need, and every action is logged.
Work with me
I work with NHS trusts, health-tech companies and research teams on design, prototypes and AI set-up. Tell me about the problem and I'll come back with a short plan and what a first prototype could look like.
Nothing you type leaves your browser and nothing is stored. Please don't share personal or health information.
