Doctoral researcher · software engineer
I research how to make adaptive AI in clinical settings trustworthy, then I build things to prove it.
PhD candidate at the University of Oulu. My research asks one question: how do you build clinical AI that keeps working safely as conditions change, and how do you prove it?
Site map
Six rooms, one researcher. Each card opens a thread of work, from doctoral focus to shipped systems and the chat that ties it together.
Your lens
Pick a perspective. The page rewrites itself to surface what matters to you.
“I'm equally at home in a research lab and a production environment. Most people have to pick one.”
Research domains
Adaptive clinical AI, runtime governance, trustworthy deployment, empirical evaluation, and a live prototype: five threads of doctoral work.
Clinical AI that learns and adapts after deployment is powerful, but also fragile. My research is about building the guardrails: systems that keep working safely even as patient populations, care protocols, and data distributions shift.
Paper I mapped the field; Paper II proposes a constraint layer over the MAPE-K loop. Paper III stress-tests whether those mechanisms actually hold when drift and adversarial feedback hit a running system.
Adaptation without accountability is just risk by another name. I design constraint layers, audit mechanisms, and rollback policies so that when clinical AI changes its behavior, it does so traceably and within defined bounds.
Four mechanisms sit over the adaptation loop: hard safety bounds, validation gates before each update, evidence anchoring to clinical guidelines, and audit/rollback when behavior diverges from what was approved.
Trust in clinical AI isn't just about accuracy. It's about whether a clinician can understand, verify, and override a recommendation, especially when the stakes are high. I work on what that actually takes to build.
That means privacy-preserving design, calibrated trust (not blind acceptance), and explainability that survives safety-critical review, not dashboards that look convincing but can't be audited.
Good ideas need rigorous methods behind them. I use Design Science Research to build artifacts, evaluate them with mixed-methods studies, and stress-test them under simulation before claims reach publication or deployment.
Simulation injects concept drift and adversarial inputs; evaluation pairs quantitative degradation metrics with qualitative clinician feedback to see whether governance holds in practice, not only in theory.
The Paper II architecture is becoming runnable code: a clinical decision support loop with MAPE-K adaptation, all four constraint mechanisms active, and an audit trail built in from day one, not bolted on after deployment.
Instantiation context is nurse-led primary-care triage. Target venue: EMSE or JSS. Comparing constrained vs. unconstrained baselines under drift is the core empirical question for 2026.
Selected work
National-scale systems, healthcare AI, and research prototypes. Hover any case file for the signal.
Sri Lanka needed a modern, scalable system for managing readmission cases across the country, replacing fragmented manual processes.
The International Organization for Migration needed a secure digital passport system for Sri Lanka with a modern web interface.
Healthcare professionals needed a secure, AI-powered tool to manage patient data and automate prescription generation.
Career arc
Five turning points that changed how I think about building AI that can actually be trusted.
“Watching a system fail and doing nothing about it isn't safety. It's observation. Systems need to act.”Read more about this shift →
“Unconstrained learning in clinical settings isn't a feature. It's a liability. Every update needs a boundary.”Read more about this shift →
“A model that's accurate but unaccountable is still a black box. Clinicians need to understand, not just trust.”Read more about this shift →
“An architecture that exists only in a paper hasn't been tested. Ideas need to run before they can be believed.”Read more about this shift →
“Making one component safe doesn't make the system safe. Governance has to be designed into the architecture, not patched in later.”Read more about this shift →
Now
What I'm investigating, building, questioning, and open to collaborating on right now.
Clinical AI adapts: it updates its models, shifts its recommendations, responds to new data. That flexibility is also its main risk. My research asks how you let a system learn without letting it learn itself into a place where it causes harm.
Paper II (accepted) proposes a Constraint Layer that sits over the MAPE-K adaptation loop. Four mechanisms keep adaptation safe: hard safety bounds, validation gates before each update, evidence anchoring to clinical guidelines, and audit/rollback when things go wrong. Paper III is now stress-testing this under simulated concept drift and adversarial inputs to see if the constraints actually hold.
“Adaptation without governance is just drift by another name.”
Experience
The path from engineering student to doctoral researcher, with national-scale systems shipped along the way.
FAST-funded PhD at the Faculty of ITEE, supervised by Nirnaya Tripathi. The research question: how do you make adaptive AI in clinical decision support trustworthy enough to actually deploy, and how do you prove it?
Master's thesis on detecting and responding to mental workload in real work settings, combining EEG signals with LLM-driven recommendations to support engineer well-being.
Full-stack engineer on national-scale government systems: e-passport infrastructure for IOM Sri Lanka, multilingual government platforms, and a readmission case management system used across the country.
Bachelor's thesis on AI-powered doctor–patient management software. Built and validated a system for managing patient records and automating prescription generation.
Publications
Research papers, working prototypes, and public presentations.
Built using Design Science Research, this architecture adds a Constraint Layer to the MAPE-K adaptation loop: four mechanisms that let clinical AI adapt without losing accountability. Evaluated with seven clinical stakeholders in a nurse-led triage setting.
A systematic review of how self-adaptive AI is being built for clinical settings. Using PRISMA and Kitchenham across 20 primary studies, it maps what's been tried, what's working, and, crucially, where the engineering and governance approaches fall short.
Can we detect when a software engineer is cognitively overwhelmed before they make mistakes? This paper uses EEG and LLMs to monitor and respond to mental overload in real work settings.
Using EEG and self-reported data to track mental workload in real time, with an LLM that suggests personalised well-being interventions based on what the signals show.
An AI-powered system for managing patient records and automating prescription generation, designed, built, and validated as part of a B.Sc. in Software Engineering.
Writing
Notes from research and engineering practice, with a focus on trustworthy AI and real-world systems.
Get in touch
Open to collaboration on medical AI, mHealth, and privacy-preserving systems. I bring both industry engineering and doctoral research to the table.
kaveebhashiofficial@gmail.com