PPE Compliance Detection Pipeline
Built a two-stage YOLOv8 object detection system: person detector on full images → cropped-person PPE detector (hard-hat, gloves, boots, vest, ppe-suit). ▪ Designed a custom bounding-box remapping pipeline to convert full-image annotations into person-crop training data. Achieved 94% mAP50 on person detection and 78% mAP50 on PPE detection. Built an inference pipeline in OpenCV for end-to-end detection and visualization.
NL2SQL Engine
Built a natural language to SQL system that uses MinHash + TF-IDF to find relevant tables before sending anything to an LLM - cut prompt size by only passing schema chunks that actually match the question, making it practical for databases with hundreds of tables. ▪ Added safety checks at multiple levels: SQL parser validates queries, table whitelisting prevents hallucinated joins, and read-only DB connections protect against write operations - plus an in-memory cache that stores query results so repeated questions don't hit the LLM again.
LLM News Pipeline
Automated an end-to-end ETL pipeline that scraped, cleaned, and structured global news articles from 20+ sources into a unified feed — reducing manual curation effort to zero. ▪ Achieved 3-class sentiment classification (positive / negative / neutral) using TextBlob with 87%+ accuracy on held-out articles, surfacing actionable tone trends across news feeds. ▪ Integrated OpenRouter LLM API via FastAPI backend to auto-generate SWOT analyses, named-entity maps, an