Prasanna Biswas

Prasanna Biswas

Advisory Research Engineer

IBM Research – India Lab
Bangalore, KA - 560071

About Me

I am an Advisory Research Engineer at IBM Research – India Lab with close to five years of experience in machine learning, deep learning, and AI systems. My work lies at the intersection of AI algorithms, system-level optimizations, and high-performance computing.

At Intel, I developed high-performance GPU kernels with SYCL for Falcon Shores architecture, optimized performance-critical operators, and explored advanced ML research combining VAEs with Diffusion Models, co-authoring papers submitted to CVR 2025 and IEEE CONNECT 2025.

Previously at Qualcomm, I optimized ONNX models for NLP, CV, and LLMs on the AI100 accelerator and contributed to custom node fusion operations for inference acceleration.

I hold a Master’s in Computer Science from IIT Bombay, where my research focused on multimodal meta-learning for sarcasm and emotion analysis.

My expertise spans deep learning for NLP, CV, and LLMs, GPU kernel optimization, AI systems, and bridging research with real-world performance.

Research

My research interests lie at the intersection of machine learning frameworks and hardware, with a strong focus on hardware-aware model optimizations. I actively study and write about GPU and ML accelerator architectures; delving into the intricacies of Tensor and CUDA cores; and the development of advanced GPU kernels that are critical for maximizing model performance.

During my time at Intel, I engaged in research generating machine-style handwriting using a diffusion-based, text-conditioned latent model combined with VAE decoding. Alongside this, I conceptualized and implemented algorithmic strategies for highly efficient compute kernels.

At Qualcomm, my work centered on compiler optimization using Graph Neural Networks (GNNs) to improve model compilation and scalability. I am thoroughly enjoying my ongoing career transition—evolving from building foundational models in academia to mastering hardware architectures, deep learning compilation stacks, and kernel development for optimal performance.

Previously, as a Research Assistant, I developed a multimodal approach to analyzing emotions in sarcasm, focusing on linguistic incongruities and the hidden sentiments behind text. My broader work in NLP has encompassed emotion classification, sentiment analysis, and sarcasm detection.

Over the years, I have maintained a deep fascination with generative modeling for conversational AI, dedicating significant effort to optimizing the computational efficiency and memory footprint of these complex models.

Some of my MTech Projects

Computational Model for Understanding Emotions in Sarcasm
  • Problem Statement: Sarcasm is a very sophisticated linguistic articulation where the sentential meaning is often disbelieved due to linguistic incongruencies. This makes emotion recognition extremely challenging.
  • Dataset Contribution: Created a benchmark dataset 'emo-UStARD' of sarcastic and non-sarcastic videos annotated with 8 primary emotions and arousal-valence levels.
  • System Contribution: Built a "One v/s Rest" classifier to predict exact emotions present in a sarcastic sentence. Proved the importance of contextual and additional information to predict emotions through rigorous experimentation.
  • Tried submissions: EMNLP 2020 Draft, NAACL 2021 Draft
Investigating Importance of Multiple Modalities in Sarcasm Detection
  • Objective: Analyze the importance of multiple modalities (like emojis) in sarcasm detection from text to better understand user mood.
  • Implemented LSTM-NN and fasttext classifier as baselines. Conducted experiments placing emojis at different text positions.
  • Analyzed effectiveness of features from SentiWord-Net and EmojiNet, achieving near 90% accuracy. Used LIME (Interpretability Tool) to confirm emojis aid in classification.
Projects for Fun

Filed Patent

Himanshu Upreti, Dheeraj Gattupalli, Vinayak Baddi, Mohit Sharma, Prasanna Biswas and Anuj Gupta. June 06, 2023.
"Pre-Processing For Deep Neural Network Compilation Using Graph Neural Networks". (Pending)

Publications

  • • Phani Kumar Nyadhsham, Prasanna Biswas, Archie Mittal. Efficient Deep Learning Model Architecture for Emergence of Machine Style Calligraphy, IEEE - CONNECT 2025. URL
  • • Phani Kumar Nyadhsham, Prasanna Biswas, Archie Mittal. Generating Machine-Style Handwriting: A Diffusion-based latent generation with VAE decoding, CVR 2025. URL
  • • Nupur Giri, Chetan Gupta, Mohit Choithwani, Prasanna Biswas, Piyush Gidwani. Home Automation Using Panoramic Image Using IoT, ICRIEECE. URL

LinkedIn Posts

I regularly write posts detailing deep dives into GPU fundamentals (like Tensor and CUDA cores), the inner workings of PyTorch's compilation stack, and other advancements in ML.

View My Posts

YouTube

Python Instructor

Playlist Thumbnail

Complete Python tutorial series from basics to advanced topics!

Watch Playlist

Technologies

Machine Learning
PyTorch torch.compile() torch-spyre
Programming
Python C++
Domain & Techniques
NLP GPU Programming Computer Vision Quantization Pruning Node Fusion Graph Optimization