Medical AI Architect (MD/PhD) | Alexander Lebedev

At a glance

Background that matters for serious clinical AI work.

Senior Medical AI Engineer (Function Health): architecture, evaluation practice, safety controls, deployment discipline
MD/PhD physician‑scientist: clinical risk, evidence, and scientific defensibility
100+ peer‑reviewed publications
$1M+ in grants and venture funding secured
16+ years across medical data, analytics, applied ML

Why clinical AI fails

Clinical AI fails when model performance is mistaken for clinical performance.

Most damage shows up after launch: workflows and data shift, quality erodes, and the system becomes hard to trace, debug, and own operationally.

The clinical job moves—protocols, coding, eligibility, and product paths diverge from what you validated; the “target” is a moving one on both sides.
Silent degradation in production—inputs and populations shift; without monitoring tied to harm, performance drops before anyone notices.
No audit trail, no incident response—you cannot reconstruct data lineage, model version, prompts/tools, and config—so failures are not explainable or fixable.
Integration is where architecture dies—latency, missing fields, upstream changes, and human-in-the-loop behavior break assumptions baked into training and evaluation.
Paper governance vs operating governance—policies exist, but there are no runbooks, change control, drift triggers, or clear accountability when something breaks.
Evaluation optimizes the wrong thing—leaderboard metrics or narrow test sets that do not match the decision, edge cases, or failure modes the product actually sees.

If you cannot explain what the system is allowed to do, how you know it is safe enough, and what happens when it is wrong, you do not have a clinical system. You have a demo.

What I help with

System architecture for clinical workflows: clear interfaces, traceability, failure containment
Safety boundaries: prohibited actions, escalation paths, auditable rationale
Evaluation: task definition, reference standards, failure‑mode testing, shift and subgroup checks, acceptance thresholds
LLM safety in healthcare: tool/retrieval boundaries, refusal policy, red‑team coverage, continuous tests
Governance that runs: change control, monitoring, incident response, post‑deployment review cadence
Executive decisions: what to ship, what to stop, what must be true before scale
Technical review for deals and boards: what is real, what is brittle, what is missing

How I work

I start from the clinical job and the harm surface. Then I design controls the team can actually run.

Typical sequence:

Use case and claims: who uses it, for what decision, under what constraints
Risk map: harm taxonomy, misuse, edge cases, “never events”
Architecture: data lineage, model boundaries, tooling, audit trails, reliability constraints
Evaluation: reference standards, test sets, failure‑mode inventory, go/no‑go criteria
Operations: monitoring signals, drift response, incident handling, change management

Typical requests

Examples of what teams bring—scoped, written, and tied to a concrete decision.

Is our evaluation defensible for launch?

Claims, datasets, failure modes, and the gaps you have not tested—before you commit in front of clinicians, customers, or regulators.

Will this architecture survive real customers, integrations, and production operations?

Data paths, versioning, reliability, monitoring hooks, and clear ownership when pipelines and products change—beyond a pilot or demo environment.

We are shipping an LLM in a health product—what breaks first?

Boundaries, tool use, escalation, red-team coverage, and continuous checks—not prompt theater.

We are investing—what is real?

Evidence versus demo, deployment realism, and the shortest list of risks that can kill the thesis—for boards, funds, and executive teams.

Selected Works

Function Health

Senior Medical AI Engineer working on clinical AI architecture and evaluation practice: safety boundaries, failure‑mode coverage, and deployment discipline in real medical workflows.

Karolinska Institutet

Research and clinical work under real constraints: traceability, defensible analysis, and execution discipline that transfers directly into evidence and governance for clinical AI.

Stavanger University Hospital

Applied ML for clinical data analysis inside institutional constraints. Secured NOK 1.5M in research funding and built collaborations that held up in practice.

Open Dialogue Space

Science Director. Built operational analytics for mental health workflows: monitoring, semantic pipelines, and decision‑relevant dashboards designed for real usage.

Dreamseer

Co-founder. Built an AI-driven wellbeing product with focus on user safety, clear system boundaries, and practical delivery.

OhCleo

Built LLM-enabled workflows for user-facing products: safe tool use, content organization, and operational reliability beyond prompt demos.

Katharsis Journeys Ltd

Co-founder. Built and operated a wellbeing business with high responsibility constraints; partnerships with academia; data/automation work that improved operations and outcomes.

Aging Research Center

Coordinated a €1M ERC-funded program and ran end-to-end clinical research execution: protocol discipline, data analysis, and defensible reporting.

Policy / public-sector analytics

Work with EUDA (formerly EMCDDA), EPA, and IIASA on evidence and analytics under public-sector constraints: traceability, defensible methods, and decisions under uncertainty.

Markets & on-chain (methods work)

Applied ML and analytics in noisy, adversarial environments: signal extraction, monitoring, and decision-making under shifting distributions.

Testimonials

Alexander is one of the most inspirational and brilliant minds I’ve encountered. A compassionate thinker who improves lives through innovation.

— Jonathan Dekle

A brilliant physician, scientist, and leader. His ability to listen, analyse, and guide with wisdom is rare.

— Tatiana Santini

Alexander brings together neuroscience, finance, behavioural economics and data to create inspiring, high-performing environments.

— Luigi Espasiano, MD

His expertise in big data analytics and statistical methods has been crucial in multiple high-impact research projects.

— Christoph Abé, PhD

He combines deep technical insight with creativity and rapid execution — nothing is impossible.

— Ivan Dimoski

A fantastic researcher and developer, warm and empathetic, with outstanding productivity and intelligence.

— Pierre de Boer

Contact

If you are building or shipping clinical AI—or reviewing it for investment—and need architecture, safety, or evaluation judgment, send a brief note.

Include:

product + clinical workflow
stage (prototype / pilot / production)
what you need to decide
where the risk concentrates (patient‑facing, clinician‑facing, regulated claims)
timeline and stakeholders

Medical AI Architect