Screening 21,379 papers with a hybrid RAG pipeline. My M.Sc. thesis was a PRISMA 2020 meta-analysis of feature-extraction methods for EEG-based sleep staging, but the harder part was the screening. I built a hybrid RAG pipeline with an LLM referee layer for content filtering, MIDO-framework structured extraction and a two-stage quality check using the Newcastle-Ottawa Scale and PROBAST+AI, narrowing 21,379 records to a 140-paper methodological pool. Applied DerSimonian-Laird random-effects meta-analysis on logit-transformed F1 scores across W, N1, N2, N3 and REM stages, quantified between-study heterogeneity and mapped feature families to stage-level performance. Published at the 4th International Sarajevo Scientific Research and Innovation Congress, 2026.
0
108
Manufacturing ERP platform with an ML layer. A cloud-hosted, multi-subdomain ERP running the day-to-day operations of a trailer manufacturer. I designed it and I am the only engineer on it: React frontend, FastAPI backend, PostgreSQL 16, nginx and Docker Compose on a VPS, plus a restorable triple-backup system. On top of that sit the models: demand forecasting, cost estimation and LLM-based document automation for quotations and technical documents, feeding daily production decisions.
0
56
Literature Harvester Suite. A production research platform that queries 13 academic sources, including PubMed, Scopus, Web of Science, IEEE Xplore, ScienceDirect, OpenAlex, Crossref, arXiv and Semantic Scholar, behind a unified provider abstraction with per-source rate limiting, retry logic, DOI recovery and health monitoring. Seven-service Docker deployment: FastAPI backend, React frontend, PostgreSQL, Redis, Celery worker and Beat scheduler, and GROBID for PDF-to-TEI parsing, with Alembic migrations, an Nginx reverse proxy, JWT authentication and encrypted database fields. Harvested 21,379 unique records and 16,453 full-text PDFs. 363 Python modules and 74 pytest suites. Sole developer.
0
48
Bosch production-line quality classifier. An end-to-end binary classifier predicting component quality failures, reducing 1,183,747 rows and 970 raw features to a modelling set of 450,519 samples and 158 features. Handled a 1:228 class imbalance with scale-weighted XGBoost and SMOTE, engineered 24 aggregate features, and raised defect recall by 29% over the baseline with threshold optimisation and stratified K-fold validation. Shipped as a containerised service: FastAPI prediction and batch endpoints, a Streamlit interface for interactive sampling, Docker Compose orchestration and SHAP-based feature attribution.
0
49
