Applied AI studio · Glasgow, UK
OhDoodle builds multimodal AI systems that read genomic data, tissue images, clinical reports and live web data together, then explain what they found in language an expert can act on.
What we do
Our own research programme. A multimodal model that reasons across histology, genomics and transcriptomics for non small cell lung cancer, designed to sit inside the NHS lung cancer pathway and support, not replace, the multidisciplinary team.
Retrieval augmented generation, document extraction, fine tuned open models and data pipelines that go into real workflows. We have shipped clinical drafting tools, compliance detection and evidence chatbots for regulated organisations.
Evaluation, guardrails and governance for AI that has to be trusted. We define the metrics before we build, test for safety and bias, and design for sovereign deployment and regulatory readiness.
Flagship research
In routine practice, DNA sequencing results, H&E pathology slides and RNA expression are reviewed in separate silos. Our platform fuses all three into a unified representation and produces a multidisciplinary team ready report per patient, with Shapley attributions and causal graphs showing why.
Validation use case: treatment recommendation in non small cell lung cancer. Training uses public, ethically approved cohorts (TCGA-LUAD, TCGA-LUSC, TCIA whole slide images), so Phase 1 needs no proprietary patient data.
Selected work
Client names are kept confidential. Pick a system to see what was built and what it changed.
Oncology technology company · London
Retrieval augmented generation paired with a 3B parameter open model fine tuned on health data. Extraction pulls key metrics from histopathology and radiology reports, a guideline layer validates them, and the system drafts a plan for the clinician to review.
International development organisation · UK
A decision support chatbot answering complex queries across more than 100 policy documents, with a fine tuned 7B open model behind it. Alongside it, document classifiers sort thousands of research articles a month for systematic review.
Property compliance platform · UK
Takes a provider's property list, scrapes live listings from room letting and short stay sites by postcode and map area, fuzzy matches addresses and titles, and labels every property as investigate or no match in a downloadable report.
Philanthropic AI programme
Links systematic review evidence on undernutrition interventions to the organisations delivering them, runs soft credibility checks on each organisation, and maps how interventions relate, so giving follows what works.
Track record
Responsible by design
High stakes decisions need AI that a regulator, a clinician and the public can all trust. Assurance is built into every system we ship, and it is a growing focus of our research.
Open weight models fine tuned and hosted on UK infrastructure, so sensitive data and model weights stay under your control. Critical decisions never hinge on a single overseas API.
Input and output guardrails, answers grounded in retrieved sources with citations, validation agents that check outputs against guidelines, and adversarial testing before release.
Model cards, data lineage, versioned evaluation reports and full audit trails, so every output can be traced back to the data and model version that produced it.
Shapley attributions and causal graphs behind every recommendation, bias and drift monitoring across patient and user groups, and a named human who signs off high stakes outputs.
Our team has built EU AI Act readiness tooling that assessed AI companies against the regulation's requirements, combining NLP, sentiment and function based metrics into downloadable compliance reports.
Approach
Every claim we make on a project is tied to a metric someone else can check. This is how a build runs.
Who acts on the output, under what time pressure, and what a wrong answer costs. For oncology that is the Day 21 MDT. For housing it is a fraud investigation.
Accuracy against guidelines, drift, causal validity, precision of flags. Baselines are run before any model is trained so improvement is real, not remembered.
Structured state, retrieval over your documents, fine tuned open models where they beat closed APIs, and validation agents that check outputs before a human sees them.
Review queues, explanations attached to every output, and dashboards that show quality over time. The system supports the expert. It does not replace them.
Process · the fourth dimension
Map the decision, the data sources and the people who own them. Agree what success looks like in one number.
Run the simplest credible approach end to end. Publish the baseline metric before any modelling starts.
Ingest documents, images or omics data into structured stores. Build retrieval and extraction that experts can inspect.
Fine tune open models or compose closed ones. Add validation agents and guideline checks around the core.
A/B against the baseline. Drift, accuracy, precision of flags. Expert review sessions on real cases.
Cloud deployment with monitoring dashboards, review queues and audit trails. Handover documentation.
Measured result against the week 0 number. Decision on scale up, with a costed plan.
Quality tracked over time. Retraining triggers, new data sources and regulatory updates folded in.
Team
Three specialists from the University of Glasgow's data science and bioinformatics programmes, with an advisory link into a leading cancer research institute.
Architecture · delivery · founder
MSc Data Science. Six years across NLP, retrieval augmented generation, fine tuning and agentic systems for health, policy and enterprise clients. Leads system design, model integration and project delivery.
Omics pipelines · validation
MSc Bioinformatics. Research experience in whole slide pathology analysis with spatial transcriptomics, and multi-omics datasets from public cancer cohorts. Leads data harmonisation and biological validation.
Training · MLOps · cloud
MSc Data Science. Healthcare AI engineering at scale, cross modal reconstruction with Vision Transformers and diffusion models, and histopathology classifiers. Leads training infrastructure and cloud deployment.
Start a project
Tell us about the decision you need to make. We will reply with a short assessment of what is feasible, what it would cost to prove, and which metric we would track.
hello@ohdoodle.co.ukOhDoodle Ltd · Glasgow, United Kingdom