Built an end-to-end audio transcription pipeline for police emergency calls, converting speech to text using ASR models and achieving high accuracy across diverse caller accents and environments. Developed an AI-powered analysis layer that evaluates call sentiment, detects escalating or critical situations, and flags high-risk interactions in real time. Implemented an automated agent performance evaluation module scoring officers on call handling quality, protocol adherence, and communication effectiveness. Exposed all functionality through a RESTful API built with FastAPI, enabling seamless integration with law enforcement dispatch and monitoring systems.

Designed and implemented an end-to-end ML pipeline on Databricks using Apache Spark for large-scale data processing and feature engineering. Leveraged Delta Lake for ACID-compliant data storage, enabling reliable versioning, time travel, and incremental data updates across the pipeline. Orchestrated model training, validation, and deployment workflows using Databricks MLflow, tracking experiments and managing the full model lifecycle.

Built a Django-based web application that extracts asset and policyholder information from insurance documents using OCR (Tesseract) and open-source vision models. Implemented an intelligent analysis layer that evaluates coverage adequacy, identifies coverage gaps, and determines whether individuals require additional or new insurance policies. Integrated an open-source LLM to generate natural-language summaries and recommendations based on extracted policy details, asset valuations, and risk profiles. Designed a structured asset registry to track user assets, map them to existing policies, and automatically flag uninsured or underinsured items.

Developed a mobile application using AI and image recognition to improve food safety for individuals with allergies. Preprocessed and annotated image datasets using Roboflow to train a YOLOv8 object detection model for identifying allergenic ingredients in food items.

Developed a local AI-powered system integrating a quantized Qwen 2.5 3B model with SQLite, MongoDB, and PostgreSQL using FastMCP. Implemented natural language interaction for data management and analysis through a unified Python client, reducing average query time by ~20% compared to manual SQL workflows.

Built a predictive AI model for wine recommendations based on food pairings using data preprocessing, feature engineering, and ML algorithms. Employed NLP and Word2Vec embeddings to match wine and food flavors with rule-based logic for pairing, achieving ~82% recommendation accuracy on test data.