Symphony

I was diagnosed with severe eczema two weeks after I was born.

My childhood was dominated by flare-ups every few days and immunosuppressive medicines that often had little or no effect. Regular appointments with dermatologists showed me what medicine could do and inspired me to pursue it as a career.

Medical training showed me the information gap from the other side. A consultation captures the flare or problem in front of the clinician, but rarely the preceding weeks: what changed, what a person tried and what made things better or worse.

Time pressure makes that context difficult to collect. The immediate problem can be treated while the longer pattern remains unclear. That gap became the product problem behind EczemaDoc.

During the Academic Foundation Programme and NHS Clinical Entrepreneur Programme, I started building EczemaDoc with my co-founders Dr Niall Jawad and Dr Noreen Akram.

The first product let people track symptoms, possible triggers and flare-ups at home. It launched on the App Store and Google Play, then became Proton Health and expanded from eczema to acne and psoriasis.

As co-founder, I worked across product definition, clinical research, fundraising, partnerships, data systems and regulatory preparation. My individual technical work included developing a multivariate flare-prediction model and designing the clinical ontology and extraction pipelines used to structure the evidence underneath Symphony.

The work repeatedly moved between patient experience, clinical constraints and operational detail. A feature was only useful if people could complete it, clinicians could interpret it and the underlying evidence remained traceable.

Proton Health joined KQ Labs, the Francis Crick Institute accelerator for data-led healthcare startups, in winter 2022. The three-month programme provided mentoring, workshops, introductions and a final demo day.

We received an Innovate UK grant in 2022 and an NIHR grant in 2023. The grants supported separate research and development work, including a smartwatch system for detecting itch and scratching. Alongside KQ Labs, we raised backing through StartUp Health and private angels to develop the product, test the commercial model and prepare for clinical partnerships.

I had spent years on strong immunosuppressive medicines. Tracking my symptoms in Proton Health found patterns that treatment alone had missed.

Two months of tracking linked strawberries, other berries and apples to my symptoms. I changed my shampoo after identifying an irritant ingredient, started using Adex Gel for its Niacinamide and trialled a handheld red-light device from Amare Skin. My last flare was in summer 2023.

That experience reinforced an important product principle: the system should help a person form better questions and observations, not present an automated conclusion as medical certainty.

Symphony is the current product. It brings together photos, symptoms, lifestyle, mental wellbeing, diet and products, looks for patterns and prepares a personalised plan. It supports eczema, acne, psoriasis, rosacea and other persistent skin concerns.

The product direction also expanded toward clinical workflows: collecting a patient’s history before a visit, summarising the important context and preparing the consultation for the clinician.

Nutrition, lifestyle and guided mental-health sources used different terms and levels of detail. Without a shared structure, records could not be compared, searched or reliably reused across conditions.

I designed a clinical ontology: fixed fields and allowed terms that describe each intervention consistently. Python pipelines extract records into that structure, validate them and write them to Airtable.

1. Define strict JSON schemas with 40 to 60 or more fields for each record type, using required fields and controlled choices to prevent vocabulary drift.
2. Model nutrition records with compounds, inflammatory markers, metabolic measures, gut-skin factors, allergens, histamine classification and dermatology outcomes.
3. Model behavioural interventions with dose, frequency, adherence threshold, washout period, contraindications, effort and rules for continuing or stopping self-observation.
4. Parse guided-session transcripts and label methods such as breathwork, cognitive reframing, somatic practice and compassion-based work.
5. Record proposed mechanisms, including vagal tone, HPA-axis activity and parasympathetic activation, alongside the symptoms or emotions targeted.
6. Split large extraction jobs into separate passes, merge only after field-level validation and map each value to the live Airtable field type.
7. Isolate errors per record and support safe reprocessing so a corrected source or schema does not create uncontrolled duplicates.

Structuring a source manually takes roughly 25 to 35 minutes. Automated extraction followed by review takes five to eight minutes, reducing handling time by about 75% while keeping failed or incomplete records visible.

I developed a multivariate model using anonymised records from 130 people. It combined symptom history, products, food, behaviour and environmental variables to predict the next day’s skin score and identify increased flare risk.

A prediction counted as correct when it fell within a predefined tolerance range of the recorded score. The leading internal test produced an F1 score above 0.80. The intended output was an earlier signal to review likely triggers—not an automated diagnosis.

Proton Health joined the Techstars Physical Health Accelerator in Fort Worth, Texas, and completed the programme with Demo Day in December 2024. We interviewed 19 dermatologists, designed the workflow with them and piloted it in three clinics. Techstars invested in the company.

That work clarified the clinical co-pilot direction: collect context before the appointment, organise it into a concise summary and give the clinician a better starting point without trying to replace their judgement.

In winter 2024, we worked with UCB to co-create the hidradenitis suppurativa portion of its app. HS is painful, long term, frequently misdiagnosed and difficult to track between appointments. Our experience with trigger data and persistent skin conditions suited that problem.

We also white-labelled the app for selected partners and worked with Gladskin, Codex Labs, Cosi Care and Eczema Clothing, connecting tracking and education with products people were already using at home.

Proton Health registered the Class I medical device with the MHRA and achieved ISO 27001 certification following an independent audit of its information-security management system.

The work covered risk management, access controls, incident processes, supplier checks and the evidence required to show that those controls were operating. Regulation was not a final badge; it changed how the company documented decisions and ran day to day.

An internal proof-of-concept analysis of 40 members compared follow-up assessment scores with each member’s baseline. Average scores improved by 34% for symptoms, 48% for mental wellbeing and 58% for people’s sense of control over their triggers.

These were product outcomes, not medically assessed trial results. As of September 2026, the current Symphony product reports more than 11,700 assessments across more than 100 doctors.

EczemaDoc began with a narrow question: could better tracking help someone understand a flare? Proton Health expanded that idea across conditions and partnerships. Symphony brought the product, evidence structures and clinical workflow together into a broader assessment platform.

The most important shift was from collecting more information to preparing better decisions. That meant designing for provenance, uncertainty, review and the different needs of a patient, researcher and clinician.

First, lived experience is valuable context but not a substitute for evidence. Second, AI and modelling are most useful when they structure information and surface patterns while keeping uncertainty visible. Third, a healthcare product is also an operating system: research, security, regulation, partnerships and commercial decisions shape what can safely reach the user.

The project connects a problem I understood personally with clinical discovery, product iteration, evidence infrastructure, applied modelling, regulation, fundraising and company operations. It demonstrates my ability to move between those layers without losing sight of the person who must ultimately use the system.