I am a full-stack data scientist working across the path from data preparation and modelling to model-lifecycle support and decision-facing applications. My work began with classical machine learning and simulation, moved through advanced-analytics demonstrations, and now centres on reliable data and ML systems for insurance. I am building deeper independent work in agentic workflows; it is a direction I am developing, not a claim of production agent ownership.

At ValueMomentum, I worked with the ERIE Insurance (US) Advanced Analytics risk platform team. My contribution was developing AWS Glue and Aurora/RDS workflows that consolidated cloud and on-premises inputs into reusable modelling datasets, and contributing across evaluation, experiment tracking, versioning, deployment, and monitoring for multiple production computer-vision applications. The wider team platform combined aerial imagery with policy and risk-location context for human-reviewed property underwriting; underwriters made the final decisions.

Earlier in ValueMomentum's Advanced Analytics team, I worked on client demonstrations in claims-data streaming and document intelligence, including LLM-assisted retrieval. These were demonstrations, not production deployments.

At Genpact, my undergraduate data-science internship covered classical ML for water-quality classification, feature analysis, Tableau reporting, and AnyLogic supply-chain scenario simulation. The simulation compared alternatives with a baseline; it is not a claim of realised savings.

Before that, during an internship with the National Institute of Disaster Management (NIDM), I worked on a small machine-learning proof of concept related to concrete-strength prediction. It was a proof of concept, not a deployed system.

In a four-person MSc team project with AstraZeneca, I led the biological-routing work while collaborating across planning and implementation. The team built research decision support for cell-line prioritisation: a white-box ranking workflow with explicit evidence-governance rules and downstream knowledge-graph / RAG explanations. Biological routes and generated explanations supplied bounded context; they could not silently change the underlying ranking. This was not a predictor of laboratory or clinical outcomes.

I am curating my earlier repositories and developing substantial, reproducible public projects that show the full path from problem framing and evaluation to data engineering, deployment, and monitoring. Public examples will use open or synthetic data and will remain separate from client systems.