user1400886

Engineered a transformation pipeline to convert raw landing content into PDF, apply metadata tagging, and securely store artifacts in Azure Data Lake Storage. Developed a text understanding and intent-detection system using few-shot prompting and retrieval-augmented generation (RAG). Automated analysis of customer service transcripts to extract key intents, issue categories, sentiment, and resolution status, converting outputs to structured JSON. Integrated an LLM-driven module to translate user questions into valid MongoDB queries with filters and aggregations. Applied prompt chaining, output parsing, and token filtering; collaborated with frontend/backend teams to expose AI APIs. Implemented data privacy measures including PII masking and secure API communications.

Human-in-the-Loop Evaluation Metric for AI Systems: Designed a metric combining log probabilities, top-k token predictions, and model uncertainty to assess AI outputs without ground truth. Built a framework to incorporate human feedback, validated across AI workflows, and documented for scalability and real-time monitoring integration.

Table Detection and Content Extraction: Developed deep learning models for table detection and integrated Azure Document AI to handle complex structures (bordered/borderless/merged/implicit tables) in noisy documents. Implemented OCR and layout analysis, edge-case handling, and pre/post-processing pipelines. Deployed containerized solutions with Docker and Kubernetes for production use.