Data Science Projects

Benchmarked 8 ML models — ANN, XGBoost, Random Forest, GBM, and more — against real surgeon estimates on 17,246 procedures. ANN achieved near-zero bias (−0.37 min) vs. surgeons' consistent −18.52 min underestimation. Published in Surgical Endoscopy 2025.

Benchmarked 5 text-encoding strategies — Label Encoding, Count Vectorization, TF-IDF, ClinicalBERT, and Sentence-BERT — on 180K surgical cases. Sentence-BERT reduced prediction error by up to 16% over traditional encodings. SSRN preprint 2025.

Context-aware SHAP explanations for deep learning anomaly detection in building energy data. Reduces explanation variability by 38% on average. Published in Energy & Buildings 2025.

Multivariate candidate-gene GWAS using MultiPhen and GATES on 2,772 elderly participants. First genomic validation of IL10 as a sarcopenia risk gene in an Iranian population. Published in J. of Biostatistics & Epidemiology 2022.

PCA and logistic regression analysis of 50 plasma metabolites (acylcarnitines + amino acids) for ASCVD risk prediction. Identified 14 significant biomarkers across four ACC/AHA risk groups. Published in Frontiers in Cardiovascular Medicine 2023.

3-arm randomized clinical trial comparing a gamified Android app (Delban) against teach-back training and usual care for post-CABG recovery. Statistical analysis via one-way ANOVA with Dunnett post-hoc tests. Published in JMIR 2021.

LLM pipeline that takes a plain-language biomedical query and identifies the most relevant human protein targets via live UniProt data, complete with 3D AlphaFold structures rendered in the browser.

Seven interactive risk modules — credit default, fraud detection, market risk (VaR), insurance analytics, and more — each powered by a real ML model on live synthetic data. Demonstrates end-to-end production risk analytics with Flask, Plotly, and REST API endpoints.

Fully interactive deep learning playground: configure and train a VAE on MNIST from scratch, then navigate its 2-D latent space with sliders, compare reconstructions side-by-side, and generate new digit images — all streamed live from a PyTorch backend.

A fully interactive 6-step educational playground: configure the dataset, build vocabulary, design architecture, watch live training, evaluate BLEU scores, and translate sentences — powered by a real PyTorch training loop streamed via SSE.

Interactive experimentation suite covering the full controlled experiment lifecycle — from pre-experiment power planning to post-experiment decision-making. Four production-quality modules with rigorous statistical methods and interactive Plotly visualizations.

A benchmarking platform that fetches real arXiv papers, indexes them with BM25 and dense vector retrieval, fuses results using Reciprocal Rank Fusion, and evaluates answer quality live. No API key required — runs entirely on open-source models.

An intelligent document analysis system using a 5-agent pipeline: Planner routes user queries, Retriever finds relevant chunks with hybrid search, Grader scores chunks in-place of LLM calls, Generator produces answers, and Critic verifies factuality. Reduces latency by 60% compared to naive RAG.

An intelligent customer support chatbot using LangGraph's ReAct architecture. Features 4 switchable free HuggingFace LLMs, 5 real dispatch tools (order lookup, FAQ search, ticket creation, etc.), real-time token streaming, and full execution transparency. Watch every agent decision unfold.