Javier Alvarez-Valle

Foundation models · Clinical AI · Somatic genomics

Javier Alvarez-Valle

I build multimodal AI that reasons about biology — from radiology and radiotherapy models used in hospitals to AI for drug discovery.

Director of AI, Quotient Therapeutics · Cambridge, UK

Generation 0 normal turns over driver divides faster protective survives injury passenger hitches a ride Somatic evolution, simulated: cells die and neighbours divide to replace them, every cell carries its own genome, and selection decides which clones take over. tap the tissue to injure it
00

From hospital images to the genome

I take AI out of the lab and into the places where it has to work — radiotherapy, radiology and, now, drug discovery.

I studied computer science at the Universidad de Oviedo and wrote my master’s thesis at the University of Cambridge. After engineering roles at Telefónica I+D, Redgate, Citrix and Featurespace, I joined Microsoft in 2012 and built Skype’s real-time telemetry pipeline, later scaled across Xbox, Bing and Office.

For 10 years at Microsoft Research I worked on AI for medicine. I led Project InnerEye — radiotherapy AI that cut contouring time by 36–67% in NHS use, open-sourced and adopted by 30+ organisations — and AI engineering for cancer treatment through FDA 510(k) clearance. As Senior Director in Health Futures I led a multidisciplinary team across continents building multimodal foundation models, with partners in three areas: radiology, deploying draft chest X-ray reports with Mayo Clinic; cancer screening with Cyted; and medical imaging in clinical trials with Novartis.

In 2026, as an advisor, I worked on autonomous chest X-ray reporting and built the knowledge-base layer of the open-source Firefly Agentic framework. Today I am Director of AI at Quotient Therapeutics, a Flagship Pioneering company using somatic genomics to discover medicines.

01

Foundation models that read medical images

My team built image encoders and report-generation models for chest X-rays — Project MAIRA — and collaborated with Mayo Clinic to deploy draft reporting.

Draft report

Every sentence points to its evidence

Grounded report generation, as in MAIRA-2: select a sentence to see the region it describes. Illustrative drawing, not patient data.

arXiv2025

Closing the Performance Gap Between AI and Radiologists in Chest X-Ray Reporting

MAIRA-X: longitudinal reporting, including lines and tubes, at near parity with radiologists on critical errors and sentence acceptability.

paper ↗

arXiv2024

MAIRA-2: Grounded Radiology Report Generation

Reports that localise every finding on the image, with RadFact to check each sentence.

paper ↗

Nature Machine Intelligence2025

Exploring scalable medical image encoders beyond text supervision

RAD-DINO: trained on images alone, it matches or beats language-supervised encoders and scales with data diversity.

paper ↗

ECCV2022

Making the Most of Text Semantics to Improve Biomedical Vision-Language Processing

BioViL: better modelling of report text substantially improves vision-language learning. My most-cited paper.

paper ↗

CVPR2023

Learning to Exploit Temporal Structure for Biomedical Vision-Language Processing

BioViL-T: learns how disease changes between studies from prior images and reports.

paper ↗

EMNLP2023

Exploring the Boundaries of GPT-4 in Radiology

GPT-4 matches or beats radiology-specific models on text-based tasks — and prompting matters.

paper ↗

More on Google Scholar ↗

02

AI in the radiotherapy planning room

Clinicians spend an hour or more outlining organs on a CT scan before radiotherapy. Project InnerEye automates the routine part — and we took it into NHS use.

Organs at risk

AI drafts the contours; clinicians review them

  • Bladder
  • Prostate
  • Rectum
  • Femoral heads

A clinician then checks and edits every structure before the treatment plan.

Organ-at-risk contouring on a pelvic CT slice, as in Project InnerEye. Illustrative drawing, not patient data.

JAMA Network Open2020

Evaluation of Deep Learning to Augment Image-Guided Radiotherapy for Head and Neck and Prostate Cancers

Up to 13× faster segmentation with machine-learning assistance, within expert variability.

paper ↗

Clinical Oncology2025

OSAIRIS: Lessons Learned From the Hospital-Based Implementation and Evaluation of an Open-Source Deep-Learning Model for Radiotherapy Image Segmentation

InnerEye in hospital use: contouring 36% faster for prostate and 67% faster for head and neck.

paper ↗

Open source2020

Project InnerEye Deep Learning Toolkit

Open-sourced so hospitals and researchers can build their own imaging models; adopted by 30+ organisations.

project ↗
03

Pathology, screening and biomarkers

The same ideas carry over to pathology, screening and biomarker discovery.

Screening

Where the model looks

  1. A whole-slide image is split into small tiles.
  2. The model scores every tile; attention shows which regions drive its decision.
  3. Clear negatives are filtered out; the rest go to a pathologist.
Weakly supervised learning on whole-slide images, as in the Barrett’s oesophagus screening work. Illustrative, not patient data.

Nature Communications2024

Enabling large-scale screening of Barrett’s esophagus using weakly supervised deep learning in histopathology

With Cyted: filtering clear negatives in oesophageal screening, cutting pathologist reviews by about half.

paper ↗

arXiv2025

NOVA: An Agentic Framework for Automated Histopathology Analysis and Discovery

An agent that turns scientific questions into executable histopathology analyses.

paper ↗

Microsoft Research2022–2025

Multimodal biomarker discovery

Foundation models and AI agents across imaging, text and clinical data.

04

AI that does the science

Agents that plan analyses, query what an organisation knows, and hand a case back to a human when unsure.

QUESTION Plan Run Check Refine Answer Ask human

Analysis agent

Plan, run, check, or ask a human

  1. Plan the analysis from the question.
  2. Run code and queries against the data.
  3. Check the evidence; refine and loop if it is not enough.
  4. Answer with a traceable record, or hand back to a human when unsure.
The loop behind tool-using agents such as NOVA and the Firefly answer agent. Illustrative.

Advisory2026

Autonomous normal chest X-ray reporting

A safety layer that reports a study as normal only when confident, and otherwise defers to a radiologist.

Open source2026

Firefly Agentic: the knowledge-base layer

Hybrid-search ingestion, schema-aware parsing, an agentic SQL retriever and an MCP server on Azure — 20+ merged PRs.

pull requests ↗

Nature Medicine2026

Evaluating the robustness and readiness of large frontier models in health AI applications

Frontier models can ace medical benchmarks without robust multimodal reasoning.

paper ↗

Microsoft2025–2026

Reinforcement learning for LLMs

Post-training frontier LLMs with reinforcement learning to improve instruction following, helpfulness and robustness.

05

Engineering that holds up in production

Before medicine I built large systems: a model only matters if it runs reliably where people use it.

Microsoft2012–2015

Skype real-time telemetry

Millions of events per second, scaled across Xbox, Bing and Office as Microsoft’s first Azure-based telemetry platform.

Microsoft Research2015

What-Dog.net

A deep-learning dog-breed recogniser that reached 3 million users in three weeks.

Nature Communications2022

Active label cleaning for improved dataset quality under resource constraints

Prioritises which labels experts should re-check, correcting up to 4× more effectively than random selection.

paper ↗

arXiv2020

Secure Medical Image Analysis with CrypTFlow

Secure multi-party inference for medical imaging models, compiled automatically from TensorFlow.

paper ↗
06

Writing

On leading teams, plus Microsoft Research blog posts written with colleagues, and interviews.

07

Recognition, talks and media

Let’s talk about AI for biology and medicine.

For roles, collaborations or speaking, email me or reach out on LinkedIn. I also mentor on MentorCruise.