I'm a PhD candidate in Computer Science at the National University of Water and Environmental Engineering. My research is about making generative AI usable in decisions that matter — where a confident wrong answer is worse than no answer.
In practice that means multi-stage LLM pipelines and how to evaluate them, multi-criteria decision analysis, and using the labour market as a live data source for questions about skills, hiring and curriculum design.
Research areas
Generative AI and large language models
Decision science and multi-criteria decision analysis
A local, fault-tolerant pipeline for evaluating candidates for high-hazard industrial roles: five-criterion MCDA, weighted Likert scores and behaviourally anchored rationales, orchestrated with Redis and Celery and served by Llama 3 8B through vLLM. Compared with GPT-4o-mini on 50 records, rank correlation was Spearman 0.745 and 78% of scores fell within one Likert point; a ten-case review found citation and weighting errors, so the method is proposed as human-supervised decision support.
Two ways to score how well a candidate matches a vacancy from the text of both: an embedding-based measure of semantic similarity, and a GPT-based pass that looks for contextual and hidden factors a cosine score would miss. Together they give a more flexible match than keyword overlap, and a path toward adaptive HR systems.
A semantic module extracts and clusters the skills employers actually ask for in job ads; a generative model turns those clusters into adaptive course structures. The point is curricula that track real market demand instead of lagging years behind it.
A career test scoring 40 factors across four groups — career growth, work environment, benefits, and personal values — matched against employer profiles built from employee feedback, so students and juniors get recommendations grounded in something other than employer branding.
Automatic categorisation of job vacancies built on managed AWS services — ECS, DynamoDB, Glue, SageMaker and Lambda — and what that architecture actually buys you in scalability and running cost.