Binsika

Built an end-to-end AI missing-person identification system. Implemented a modular pipeline using OpenCV for feature extraction, SQLite BLOB embedding storage with composite indexing, and cosine-similarity search. Optimized query latency to under 15 seconds and achieved 70–80% precision on 50–100 face gallery entries via a modular Flask interface. Developed a Python tool to parse source code into an AST and dump into a custom JSON format for LLM checks to catch bad patterns and logic bugs. Split the system into modular Flask services to cache responses and reduce analysis time from ~8 seconds to under 3 seconds. Also engineered a predictive maintenance system for factory vehicle sensors to detect failing parts before breakdown. Designed custom physics-inspired features (e.g., heat deltas and tool wear ratios) and trained models (Random Forest and XGBoost) with comparisons on heavily skewed data to find approaches that identify real failures without spamming false alerts.