Project - 'Efficient Particle Transformer for Jet Tagging' Jan 2024 – Jun 2024
Project - ‘Acceleration of ray tracing algorithm for personalized audio’ Jul 2023 – Present
• Accelerated ray tracing algorithm with CUDA for real-time processing for immersive audio experience.
• Employed locally weighted regression to find Fourier transforms of material filters.
• Achieved personalized HRTFs (head related transfer function) with loss < 6-10 dB from ground truths.
Project - ‘Optimization & acceleration of Transformer-based models’ Apr 2023 – Present
• Achieved 2x speed up with ~60% compressed Transformer (BERT) based models on GLUE dataset.
• Implemented masked pruning, additive powers-of-2 quantization with accuracy degradation <1%.
• Compressed Attention module with SVD & PCA by finding optimal low-rank for each layer.
Project - ‘Resource-constrained AI Deployment on edge’ Jun 2021 – Aug 2021
• Built efficient DL models to predict constant state of motor at edge with 5x reduction in memory footprint.
• Implemented state-of-art ARM CMSIS-NN kernels to achieve 80% accuracy on edge.
• Minimized accuracy loss to 0.7% & deployed model on a Nordic microcontroller with ~256 KB RAM.
Project - 'Robotic Process Automation' Feb 2021 – Apr 2021
• Saved 100+ hours of monthly manual labour by automating data extraction from invoices using OCR & vision.
• Integrated services - automatic form filling, speech-to-text transcription & diarization with RPA web platform.