Taha Ferhan

• Building a spatial mixed logit model of consumer spending behavior across retail outlets to estimate revenue, risk, and support expansion decisions at a granular level. • Modeling heterogeneous consumer tastes from large-scale transaction and geospatial data, capturing how geographic frictions shape spending patterns and competitive interactions through cannibalization effects across locations in Morocco, yielding demand elasticities across regions. • Building scalable geospatial data pipelines on integrating satellite imagery, POIs, and mobility data into a unified analytical layer. • Scaled inference to 1M+ locations and 1,900+ stores using GPU computing, addressing endogeneity to improve identification and producing outputs used for location planning and decision visualization.

• Integration of satellite imagery analysis into expansion decision workflows. • Designed and trained a Mask2Former semantic segmentation model on AWS GPUs to classify Moroccan building typologies (villa, apartment, traditional housing) at ~0.3 m resolution, generating spatial proxies for population density and purchasing power. • Developed an end-to-end dataset pipeline: extracted satellite tiles and OSM building footprints, processed imagery with GDAL, converted annotations to COCO format, and curated labels in CVAT for model training. • Deployed the trained model as a QGIS plugin, enabling geospatial inference and visualization directly within analyst workflows.

• Developed an intelligent chatbot for data.gov.ma to assist users with general inquiries and dataset searches. • Implemented three transformer-based models: sentence similarity, intent classification, and translation, along with the spaCy French corpus model. • Fine-tuned CamemBERT to classify user intents (dataset queries vs. general questions) using generated data, available in my Hugging Face repository (tferhan). • Utilized FAISS to build a search index for identifying similar words in user queries, enhancing dataset recommendations. • Deployed the backend on an Ubuntu cloud instance using FastAPI and Docker, providing a scalable and secure hosting environment. • Developed the chatbot interface with JavaScript, HTML, and CSS, and integrated it into Drupal through a custom plugin built in PHP.