Shreyas Dixit

Lead the development of end-to-end AI systems for Embodied AI, integrating computer vision, kinematics, and real-time inference backends.

Designed and implemented an AI pipeline for a dexterous robotic arm achieving 97% task accuracy and 2mm precision for object manipulation tasks under challenging conditions like fog and occlusions.

Developed the AI-Hand Bartender system, enabling autonomous serving through sequence learning and object detection, and created a novel video-based learning mechanism to extract 6D pose data from human demonstrations.

Explored reinforcement learning (RL) in NVIDIA Isaac Sim using algorithms like ALOHA and Diffusion Policy, and implemented real-time camera calibration requiring only six examples.

Engaged in research on computer vision, AI detectability, and LLM hallucination, focusing on watermarking techniques and provenance layers.

Developing "Peccavi", a visual paraphrase attack-safe watermarking technique to ensure traceability and authenticity of synthetic media.

Served as Associate Organizer of the Defactify 4.0 workshop at AAAI 2025, under the guidance of Dr. Amitava Das to enhance AI robustness.

Developed a Stutter Detectionsystem utilizing Wav2Vec2 and Agnostic BERT to detect stuttering timestamps and compute percentages.

Trained models on Indian datasets achieving a WER of 0.31, and built an accessible front-end application for on-device deployment.

Led the Microsoft Learn Student Club at VIIT Pune, fostering a strong technical learning community.

Mentored students and organized machine learning initiatives (Nov 2022 – Aug 2024).

Built an end-to-end system using open-source deep learning models to generate contextually relevant titles for millions of Instagram short-form videos.

Researched, optimized, and deployed the system to production with strong cost and time efficiency.