Shibin Thomas

• Built a data extraction workflow that converts unstructured supplier descriptions into structured component parameters using internal records, external APIs and technical data sheets (Web Agent), with human review for uncertain cases. • Evaluated two settings: single-category test: 97.5% accuracy on 142 ceramic-capacitor records; multi-category test: field-level F1 increased from 46.0% to 78.2% across 133 records, 825 annotated fields and 15 component categories.

• Improved multi-turn tool orchestration, reducing end-to-end latency by 30% and increasing query-resolution success by 20% using planning and reasoning. • Developed a low-latency multimodal Voice AI assistant integrating GPT-Realtime voice capabilities, conversational memory, and guardrail-based safety controls. • Designed a LiteLLM-based inference layer with routing and fallback across hosted and local LLMs, supporting reliable operation across multiple providers.

Collaborated with clients and the R&D team, resulting in a 20% improvement in overall process efficiency. Collaborated closely with business experts and stakeholders to implement advanced AI and NLP solutions, significantly enhancing the functionality of products for 2 key clients. Worked on implementing a chatbot using NLP technology in Appian, improving user interaction and support response time by 30% for better customer satisfaction.

Collected and processed satellite data for ML model to improved crop health monitoring by 20%. Utilized CNN models for image analysis, enhancing the platform's ability to predict crop diseases with over 85% accuracy