Project 1: MedseneAI — OCR-Based Medical Document Automation (Python, Flask, MySQL, PyTorch, Scikit-learn, OpenCV, Tesseract)
● Designed an OCR document classification engine to identify page-level formats (Check, EOB, VCC) using TF-IDF vectorization and cosine similarity scoring with threshold-based selection and recursive fallback logic, contributing to a 27% improvement in field-extraction accuracy.
● Built a template detection stage using fuzzy matching (SequenceMatcher / Levenshtein ratio) and geometric (X,Y) coordinate validation to identify the exact layout template for accurate field extraction.
● Built a TF-IDF + SVM field validation model to verify extracted values map to the correct fields (e.g., Check Amount vs. Patient Name), with confidence scoring to flag uncertain extractions.
● Developed a fuzzy matching module using prefix-based candidate filtering and similarity thresholds to correct OCR spelling errors, improving label recognition and database lookup accuracy.
● Converted scattered OCR word output into structured, readable lines using alignment logic (top position + left order), enabling accurate extraction of multi-line values like addresses and descriptions — this line-reconstruction step drove the 33% reduction in manual document sorting.
● Corrected page rotations (90°, 180°, 270°) using Tesseract OSD and fixed slight skew using Hough Line Detection; improved processing speed for large multi-page documents using multithreading.
● Created Joblib model files for domain-specific keywords (Check No, Patient Name, Amount) enabling keyword-based field extraction at document ingestion.
Project 2: SweetHello AI — AI-Powered Patient Call Assistant (Python, Flask, REST APIs)
● Built an AI call-handling flow that automatically answers inbound patient calls and engages in spoken dialogue to resolve common queries without human intervention.
● Implemented smart escalation logic that transitions calls from AI mode to a live agent based on patient responses and issue complexity.
● Integrated the AI dialogue system with backend REST APIs to fetch patient-specific information and provide context-aware responses during live calls.
