Three years of production analytics experience: building ETL pipelines, automating reporting workflows, and shipping dashboards that teams actually use. Currently deepening my ML skills in my master's — from fairness-aware modeling to computer vision.
DMA
The stack I work in daily, from data engineering to modeling to the last-mile dashboard.
Academic and personal work from my M.Sc. — each links to a repository with the full notebook or code.
Replicated bias-mitigation experiments from arXiv:2412.09896 end-to-end on the Adult Income dataset — pre-, in-, and post-processing interventions, with post-processing cutting demographic parity difference by 99.7% at a 2.2pp accuracy cost.
View repositorySemi-supervised classification of code-switched English/Roman-Urdu tweets into 6 categories, reaching up to 92% accuracy — my B.Sc. final-year project with a co-authored research paper.
View repositoryBenchmarked 5 regressors on 1M vehicle listings — gradient boosting and an entity-embedding neural network reach ~4–5% MAPE where linear regression fails at 33%.
View repositoryReal-time crypto analytics on a decoupled architecture: a FastAPI backend shielding a Streamlit/Plotly frontend from the CoinGecko API, with caching and rate-limit handling.
View repositoryExploring a 4-D global climate reanalysis grid (37 pressure levels, 0.25° resolution) — memory-safe NetCDF handling, interpolation comparisons, and multi-scale visualization.
View repositoryEnd-to-end image classification packaged as CLI tools — data validation, preprocessing, and a Keras CNN with leakage-safe class balancing.
View repositoryResearch project on artifact-free deep-zoom fractal rendering: a plugin-based coloring architecture with C0/C1 continuous iteration counts.
View repositoryEnd-to-end cloud ML pipeline: EC2 training, serverless Lambda inference, S3 and DynamoDB storage, and load-balanced serving behind an ELB.
View repositoryThe four dashboard designs my clients relied on most — rebuilt from scratch with sample data so you can click, filter, and cross-highlight them live, exactly as they run in production.