// PROJECTS
FEATURED WORK.
A selection of projects from work and personal exploration. Each one had a real problem to solve.
CreditWatch AI
Built XGBoost and logistic regression models on 70K+ account records for 90-day default prediction — 18% better than the legacy bureau scorecard. Engineered 45+ features from transaction history lifting AUC by 11%. SHAP explainability dashboard adopted by credit analysts in weekly risk reviews. Currently re-architecting into a multi-agent LangGraph system for automated risk summaries.
Healthcare OCR Pipeline
Built a document ingestion pipeline using OCR and SpaCy NER to extract structured data from paper-based provider enrollment forms — cutting manual review effort for the claims team by 40%. ETL pipelines fed ARIMA forecasts into ServiceNow to auto-flag bottlenecks, saving support teams 6–8 hours per week.
Multi-Model Video Understanding
End-to-end system that converts raw video into structured, searchable metadata. YOLOv5 for object tracking, DeepFace for identity resolution across scenes, VOSK for speech-to-text with a BiLSTM for speaker segmentation. BLIP fine-tuned on Flickr30k for frame-level captions — BLEU score improved 40% over the base model.
Argus Agent System
AI-powered DevOps monitoring agent that watches Jenkins pipelines, detects anomalies with ML, and surfaces LLM-generated insights. The argus-cortex layer is a self-improving meta-agent that audits and rewrites the core agent's monitoring logic using a Planner / Generator / Evaluator pattern.