Projects

—

تست مسیر انتشار

—

Neural nets on 7,200 thyroid screening records: the most accurate model misses the most patients. PyTorch, focal loss, AUPRC, cost-vs-threshold analysis.

An operations-intelligence stack on 99,441 real e-commerce orders: star schema, an 18-configuration forecasting bake-off (GRU, LSTM, LightGBM vs statistical baselines), inventory cost simulation with bootstrap confidence intervals, revenue forensics, and a governed natural-language analytics agent.

PPE compliance checker using classical HSV color segmentation and contour detection: helmet/vest localization, IoU, and a three-level evaluation (detection, localization, compliance decision).

Visual QC classifier for cast/machined metal surfaces (good vs defect, subtype-tagged) using HOG + LBP features and classical ML, evaluated with per-subtype accuracy and ROC-AUC.

Tool-using conversational agent with memory: TF-IDF routing, doc retrieval, and pronoun coreference resolution, evaluated on held-out routing and multi-turn coreference accuracy.

TF-IDF + Logistic Regression financial headline sentiment classifier with a random-vs-held-out-template evaluation comparison, plus a sentiment-to-next-day-return trading signal demo.

Hybrid RAG (BM25 + LSA via Reciprocal Rank Fusion) over an enterprise knowledge base, with a real retrieval evaluation (Recall@k, MRR@k) against 40 labeled questions and cited extractive answer synthesis.

Markowitz portfolio optimization (Ledoit-Wolf shrinkage covariance, efficient frontier, risk parity) with a walk-forward out-of-sample backtest across four allocation strategies.

XGBoost fraud classifier over 100K transactions with a time-based split, precision/recall operating points, and a simulated real-time scoring latency benchmark.