Geosite accessibility & discovery: identified and fixed target leakage in an initial national ML pipeline, then rebuilt the study on 1,662 labeled sites across Morocco's 12 regions, integrating Copernicus DEM, ESA WorldCover, and OSM data

Ran 500m-clustered cross-validation with McNemar significance testing across four tree-based models; final model reached 78.6% / 73.5% national accuracy, with a complementary presence-background model reaching 0.956 AUC to flag potentially undiscovered sites

Cr(VI) adsorption modeling: modeled chromium(VI) adsorption onto iron-grafted biochars from 205 experiments across 15 literature studies, using study-grouped and leave-one-study-out validation with nested Optuna hyperparameter optimization

Random Forest model achieved R² = 0.893, validated on a completely unseen material to stress-test generalization under real distribution shift

Built the platform's AI pipeline from scratch: a FastAPI microservice integrating Whisper transcription with LLM-based coaching and pitch generation for video-CV analysis, deployed on AWS ECS and integrated with a Laravel backend

Designed injection-resistant prompts using role/data separation and delimiters for untrusted input, with LLM-based content-validation gates rejecting insufficient or off-topic submissions before the output stage

Operated AWS infrastructure (ECS, RDS, IAM) supporting reliable production operation of the AI service

Implemented adaptive learning mechanics in Unity using TARL methodology, improving student engagement and retention in prototype tests

Contributed to the TARL ecosystem (Parent App, Teacher Dashboard, Final Test Platform) as part of a 5-person Agile team