Bart Amin

I developed a fine-tuned retrieval head for RAG that learns to more reliably retrieve relevant passages by transforming the query embeddings before retrieval. It is trained on synthetically generated question-chunk pairs from the corpus, and benchmarked against standard top-K cosine similarity on the isaacus/legal-rag-bench legal corpus.

CDRAG is a new retrieval framework that uses hierarchical document clustering, and LLM-guided document selection from those clusters to curate the context most efficiently. It is benchmarked against standard top-K RAG on 100 legal questions from the Legal RAG Bench dataset.

Compares centroid, hierarchical, distribution, and density-based clustering algorithms (K-Means, AHC, GMM, OPTICS, HDBSCAN) across multiple datasets with visualization and cluster quality metrics in R.

Benchmarks Time-Varying VAR models (TV-VAR GAM and LTV-VAR) on their ability to detect non-stationarity in multivariate time series, with ROC and Type I error analysis.