Alejandro Sanchez Guinea

Scientific direction for deployed AI

Chief Scientist at EyeTrustAI

I created and lead EyeTrustAI’s research programme on risk control for AI systems that act, adapt and operate under changing conditions.

From foundational questions to reproducible evidence and product architecture.

Visit EyeTrustAI Explore the public research
Research directionDefine the scientific programme and priorities.
Methods and evidenceDirect theory, experiments and reproducibility.
Product translationTurn research into risk-control primitives.

Research programme

Closed-loop conformal risk control

The programme advances conformal risk control beyond static calibration and isolated predictions toward adaptive, deployed systems.

Track I

Certificates across changing environments

Where is a risk certificate valid?

Transfer and maintain statistical guarantees when deployment context changes.

  • Contextual selective risk control
  • Meta-conformal transfer
  • Certificate expiry under drift
  • Adaptive risk contracts
Track II

Control inside closed-loop AI systems

What happens once the certificate controls the system?

Study sequential action, selective feedback and policies that alter future data.

  • Sequential conformal authorisation
  • Policy-dependent feedback
  • Safe active auditing
  • Performative risk control
Where the tracks converge Can a risk certificate transfer, act, receive feedback, adapt and remain statistically meaningful?

Featured research advance

Sequential safety for clinical decision pathways

The work extends conformal prediction from isolated image outputs to multi-stage workflows in which diagnostic actions reveal information and shape later decisions.

View related reproducible research
Scientific contribution

Pathway-level guarantees

The guarantee is pathway-level: at the target coverage, at least one guideline-consistent safe pathway remains reachable.

Decision structure

Uncertainty across stages

A shared miscoverage budget controls pruning as tests reveal information and the pathway evolves.

Empirical scope

Five public datasets

Ophthalmology studies examine pathway retention, decision risk and the operational effect of allocating uncertainty across stages.

Public research evidence

Validation across domains

Public repositories make the methods, experiments, results and limitations inspectable.

Medical image classification

Prediction-set usefulness

Calibration support, reference labels and expert disagreement across four imaging studies.

View repository ↗ Medical image segmentation

Omission risk and transfer

Expert disagreement, region size and the cost of recalibration across retinal datasets.

View repository ↗ Incident analysis

Calibration over time

Evidence requirements, controller ageing and recalibration in aviation and consumer-injury data.

View repository ↗ Social media analysis

Automation under error budgets

Emotion, sentiment and crisis-triage policies that trade automatic action against human review.

View repository ↗

Research translated into systems

From certificates to operational control

Changing environments

Shift-Aware Risk Control

Global, context-transfer and context-weighted methods with explicit validity, transfer and leakage audits.

View repository ↗
Agent action control

CodingActionGate

A research prototype for inspecting proposed actions and routing them to proceed, defer, escalate or block.

View repository ↗

EyeTrustAI

The research programme informs the Risk Gate and its deployment model.

Visit EyeTrustAI Explore the broader research