I work on systems that adapt within the limits of their own competence and hand control back to a person at the boundary of it — how they should be elicited, architected and evaluated, particularly when the intended users are people the default design excludes.
The position emerged independently across three projects in three domains. My interest is in how such systems should be elicited, architected and evaluated, particularly when the intended users are people the default design excludes.
An autonomous transport service whose decisions escalate to a human at the boundary of what the system can safely handle.
Automation operates only where the enabling infrastructure exists, and the household retains control where it does not.
Non-urgent cases are routed automatically; the system defers to a nurse line or emergency services at the boundary of what it can safely handle.
Currently completing an MSc in Computer Science at Malmö University — 60 advanced-level credits, four A grades and one B in the first year, class representative for the programme. Thesis year 2026–27. Seeking a funded PhD position.
My path to computer science research runs through two other degrees and fourteen years outside a university. I studied Business Marketing and Management at Lebanese University before spending over a decade directing digital transformation work across international markets, then returned to formal study in 2022 — first for a BSc in Computer Science, now an MSc at Malmö University focused on the systems research below. The route was long, but it is why the exclusion and adoption side of the work is empirical rather than theoretical.
180 credits over three years on the Software Development track, degree completed in 2025. The degree project — a grid-based approach to parsing 2D web interfaces — was graded Pass with distinction (5). It followed an earlier degree in Business Marketing and Management at Lebanese University, roughly 180 ECTS, which is where the business and organisational half of the research interest started.
Before returning to university, I spent over a decade directing digital transformation for international media brands — leading cross-functional teams across four countries, and seeing first-hand which systems people actually adopted and which ones they quietly worked around. That gap, between what a system is designed to do and what excluded users actually do with it, is the empirical starting point for the research position above.
Built a pipeline to test whether spatial and numerical relationships between UI elements, independent of high-level object recognition, could support automated parsing of 2D web screens. Python pipeline using Selenium and ChromeDriver for controlled capture, OpenCV preprocessing, YOLOv8 and Tesseract OCR for detection, and a rule-based agent layer that tested whether parsed elements were actually clickable, so the score reflected interaction feasibility rather than layout structure alone. Evaluated on 290 screenshots across e-commerce and media domains. Two of five hypotheses failed. Regressing file-based compression on variability times density returned R² = 0.0002, showing that visual compressibility is largely orthogonal to spatial and categorical complexity — a counter-intuitive result the examiner identified as the basis for continued work on screen perception that is not grounded in high-level objects. 8×8 grids performed best overall, with grid consistency reaching 96.3% on e-commerce layouts and 92.1% on media layouts. Composite weights were selected by testing five configurations against blind rankings from three reviewers, matching human judgement on 85% of the sample.
Sole software engineer on a two-person project team designing an autonomous special transport service for elderly users and people with disabilities, addressing the cost, availability and stigma limitations of existing services such as Färdtjänst. Responsible for the software architecture and for building a scenario-based simulation environment used to evaluate the service with participants. The work was grounded in participatory methods run by the team: literature review, co-creation workshops, and interviews with users, service providers and caregivers. Project internals are under NDA.
Asked how AI and IoT household energy optimisation can reduce energy poverty risk while treating thermal comfort as a hard constraint rather than a variable to be traded away. Reviewed four literature strands: comfort-constrained optimisation of residential energy systems, low-burden energy management that minimises behavioural demand on the household, coaching-based energy-poverty interventions, and large-scale field evaluations of smart home deployment. Identified the gap that technical optimisation research and energy-poverty intervention research proceed largely in isolation, and proposed a three-layer architecture separating sensing and control, prediction and optimisation, and human coaching with preference constraints and manual override, so that automation operates only where enabling infrastructure exists and the household retains control where it does not. Positioned explicitly as a design-level synthesis from literature review and qualitative comparative analysis, not a validated system.
Conducted two semi-structured interviews with newcomers in their first year in Sweden, using a behaviour-first protocol that excluded hypotheticals and leading questions. Bottom-up affinity analysis and interpretive lenses for workarounds, contradictions and emotional strain reframed the problem from digital exclusion to a first-contact gap: knowing the correct healthcare channel does not mean being able to use it at the moment of need, given e-ID, language and prerequisite barriers. Designed a minimum viable experiment with hypotheses and decision thresholds set in advance, then ran four scripted scenarios against the prototype, including an emergency case verifying that urgent routing suppressed all commercial prompts. The runs validated service logic and left the behavioural assumptions untested, which was the finding. The escalation hypothesis failed on measurement rather than result: the design had not distinguished a visible nurse-line option from a tapped one, making the construct unmeasurable as specified. Documented the fix alongside the study's biases, including untested classifier performance on Arabic, Somali and Tigrinya symptom descriptions and the tendency of fake-door flows to overstate willingness to pay among users under financial pressure.
View live prototype →A global Arabic oldies radio network built, launched and operated single-handedly — streaming infrastructure, four independent feeds, native Android and iOS applications in Flutter, music curation across the 1980s, 1990s and 2000s, and station jingle production. Every layer of the product is my own work, from server configuration and app development through to programming decisions and audio production. In continuous operation since launch, with an audience concentrated across diaspora communities and Arabic-speaking home markets.
Four feeds serve distinct listener profiles, from loyal daily tuners to broad discovery audiences. The fastest-growing feed reached 153 countries with 85% growth over a thirty-day window, and the engagement-focused feed sustains an average session length well beyond typical digital media interaction.
Distributed pipeline over a 9GB Amazon review corpus, roughly 20.8 million rows after cleaning, using Hadoop HDFS and PySpark. Owned the feature engineering stage: rating-derived sentiment labelling and keyword-based platform categorisation. The dataset exceeded available memory, so completing the pipeline required tuning Spark driver and executor allocation, memory overhead and network timeout after repeated crashes. Descriptive analysis without inferential testing; sentiment is derived from ratings and platform is inferred from product keywords, so results describe review populations rather than users.
The research position depends on being able to build the systems it studies, not just describe them. I pair the methods below with the engineering to prototype, instrument and evaluate systems directly — including AI-assisted development as an established part of the workflow, from pipeline scaffolding to literature synthesis.
Innovation for Change in a Digital Society. 60 advanced-level credits completed. Four A grades and one B in the first year. Class representative for the programme. Thesis year 2026–27.
Vertically Integrated Projects Programme · Sustainable Digitalisation Research Centre.
180 credits, degree completed. Degree project graded Pass with distinction (5).
Mathematics Fundamentals (MA101C), Fundamental Programming (DA110I), Introduction to Computer Science (DA100D).
Global Arabic oldies radio network: streaming infrastructure, four feeds, Flutter applications for Android and iOS, music curation and jingle production. 81,600+ monthly listeners across 169 countries.
Approximately 180 ECTS.
I'm seeking a funded PhD position in self-adaptive systems, human-in-the-loop architectures, and the design and evaluation of systems for users the default design excludes. I'm glad to hear from research groups, potential supervisors, and collaborators working on adjacent questions.
Full working rights in Sweden — no sponsorship required. Based in Eslöv, Skåne.
Arabic (native) · English (professional) · Swedish (beginner)