Dhiya Devassy

Fine-tuned LLMs via RLHF and SFT on a dataset of 10,000+ instruction-response pairs, improving response quality by 25% while boosting accuracy. Implemented targeted LLM alignment strategies to reduce model hallucination rates by 30%, collaborating with cross-functional peers to enhance system safety for 500+ daily production users. Optimized pipeline inference latency by 15% through precise prompt engineering and efficient token management solutions.

LLM Post Training Intern - Green Rider Technology LLP - Remote

(2026-03 - 2026-05)

• Fine-tuned LLMs via RLHF and SFT on a dataset of 10,000+ instruction-response pairs, improving response quality by 25% while boosting accuracy.
• Implemented targeted LLM alignment strategies to reduce model hallucination rates by 30%, collaborating with cross-functional peers to enhance system safety for 500+ daily production users.
• Optimized pipeline inference latency by 15% through precise prompt engineering and efficient token management solutions.

Cyber Security Intern - Kerala Police Academy (KEPA) - Thrissur, Kerala

(2025-05 - 2025-07)

• Engineered an intelligent AI-based Facial Recognition System using OpenCV and FaceNet, increasing verification accuracy to 92% across a database of 1,000+ suspect profiles.
• Streamlined criminal real-time person identification workflows, successfully reducing manual investigation timelines by 40% for active cybercrime caseworkers.
• Integrated OSINT intelligence tools with computer vision pipelines, improving efficiency by 3x through structured problem identification and strategic data mapping.

Bachelor of Technology - Artificial Intelligence and Data Science - Jyothi Engineering College, APJ Abdul Kalam Technological University (2022-09 - 2026-04)