Sowmiya Devi S

[Horizontal Integrations pvt Ltd]

• Developed an advanced AI chatbot to support HVAC operations within the Tracer Ensemble BMS using Azure OpenAI GPT-4, LangChain RAG, and ChromaDB for contextual document retrieval and dynamic response generation. Integrated Flask for backend processing and Streamlit for the front end, with cloud deployment planned for enterprise-scale access. Enhanced operational support by automating complex query handling and document referencing.
• Engineered a proof-of-concept system for automated HVAC duct network detection and zone mapping from floor plans. Leveraged YOLO for VAV and diffuser detection and applied the Segment Anything Model (SAM) for precise room segmentation. Designed a hybrid pipeline incorporating Shapely geometry for boundary refinement based on wall constraints, achieving 75% accuracy in zone clarity and significantly accelerating HVAC design workflows.
• Developed an automated multimodal labeling pipeline for 300+ HVAC equipment types in Tracer-TGE, combining Vision Transformer (ViT), ResNet50, and YOLOv8x for high-precision component recognition in AHU equipment category. Integrated static AHU diagrams with dynamic GIF animations to visualize real-time equipment states (e.g., fan operations, damper positions). Enabled XML-based configuration uploads, reducing manual labeling time while sustaining over 90% accuracy, thereby optimizing HVAC documentation processes.

• Designed and deployed an LLM-powered chatbot to automate company policy inquiries regarding HR, leave, and insurance claims. Utilized Retrieval Augmented Generation (RAG) with LLaMA-2 for natural language understanding and accurate response delivery. Developed a full pipeline from data collection and chunking to embedding in a vector store, improving internal policy query resolution and employee self-service efficiency.
• Built a robust NLP pipeline for detecting Protected Health Information (PHI) and Personally Identifiable Information (PII) in healthcare communications. Combined BERT-based NER models, BIO tagging, and regex methods to enhance entity detection accuracy. Designed an automated email alert plugin that proactively flags potential PHI leaks, reinforcing compliance with healthcare data privacy regulations.
• Developed an end-to-end OCR solution for extracting data from government ID cards (Aadhar, Passport, Visa, DL, Voter-ID, Emirates ID). Employed a multi-stage pipeline: Faster R-CNN for image segmentation, Inception-ResNet-v2 for image orientation correction, Custom PaddleOCR models (DBMobilenetV3 for detection, SVTR_LCNET for recognition). Utilized SpaCy NER and regex for post-processing and entity extraction. Delivered structured outputs with bounding box coordinates for downstream processing, improving customer onboarding accuracy in retail verification workflows.

• Led development and analysis efforts for banking projects, ensuring timely delivery and adherence to quality standards. Mentored and guided three junior developers, facilitating their growth and development within the team.
• Recognized as 'Rock Star Rookie' in Q4 2019 for outstanding contributions to project success and team collaboration.