About me | Research | Patent | Publications | Youtube | Projects | Technologies I am currently working as an AI Software Solutions Engineer at Intel Habana, Bangalore, where I specialize in developing high-performance kernels for Intel’s next-generation GPUs. My work involves optimizing latency, memory bandwidth, I/O access, and compute utilization using SYCL, with a particular focus on enabling dynamic shape support for advanced workloads.
I have co-authored two research papers, one submitted to CVR 2025 and another to IEEE CONNECT 2025.
Prior to joining Intel, I worked as a Senior Machine Learning Engineer at Qualcomm, where I was responsible for optimizing ONNX models for Natural Language Processing (NLP), Computer Vision (CV), and Large Language Models (LLMs) to enhance inference speed on Qualcomm’s AI100 accelerator. My role involved designing and implementing software modules for AI/Deep Neural Network frameworks using C++ and Python. Additionally, I contributed to the implementation of node fusion as a custom operation for performance optimization.
Before my tenure at Qualcomm, I was a Research Assistant at the Center for Indian Language Technology (CFILT), Indian Institute of Technology Bombay, where I was part of the IBM-AIHN Network. My research focused on analyzing the impact of sarcasm on emotion detection using a multi-modal approach. My primary area of interest is Machine Learning for Natural Language Processing (NLP), with a strong passion for language generation, language models, and multimodal NLP.
At Intel, I was involved in research on generating machine-style handwriting using a diffusion-based text-conditioned latent model combined with VAE decoding.
Previously, as a Research Assistant, I developed a multimodal approach to understanding emotions in sarcasm, focusing on the incongruities in language and the hidden emotions behind textual expressions. My work in text-to-text NLP has revolved around emotion classification, sentiment analysis, and sarcasm detection.
Over the years, I have been particularly fascinated by generative modeling for conversational AI and have dedicated significant time to optimizing computational efficiency and memory requirements for such models. Additionally, I have worked with Graph Neural Networks (GNNs) for efficient compilation of ML models, enabling improved model performance and scalability.

1. 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)

• Phani Kumar Nyadhsham, Prasanna Biswas, Archie Mittal. Efficient Deep Learning Model Architecture for Emergence of Machine Style Calligraphy (Submitted to IEEE CONNECT-2025).
• Phani Kumar Nyadhsham, Prasanna Biswas, Archie Mittal. Generating Machine-Style Handwriting: A Diffusion-based latent generation with VAE decoding (Accepted at CVR-2025).
• Nupur Giri, Chetan Gupta, Mohit Choithwani, Prasanna Biswas, Piyush Gidwani. Home Automation Using Panoramic Image Using IoT, International Conference on Recent Innovations in Electrical, Electronics & Communication Engineering (ICRIEECE). URL: Home Automation Using Panoramic Image Using IoT

Python Instructor

---

Projects for fun

1. K-Means clustering from scratch: K-Means-Clustering
2. GPT from scratch - by Andrej: NanoGPT
3. Term-Deposit Prediction: Term-Deposit-Subscription-Prediction-via-PySpark
4. Classic Snake Game: Classic-Snake-Game
5. Coding Questions: Questions for fun

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 the linguistic incongruencies or differences in implied and surface sentiment. This makes the emotion recognition task extremely challenging in case of sarcasm.
• We had two major contributions on this project:
• Dataset Contribution:
• Previous works have extensively studied sentiment and emotion in language, while the relationship between sarcasm and emotion has been largely unaddressed. To address this, we created a benchmark dataset 'emo-UStARD', of sarcastic and non-sarcastic videos, that is annotated with 8 primary emotions, and also arousal - valence levels to get the intensity of emotions.
• System Contribution:
• Sarcasm is generally associated with negative emotion. The question is which negative emotion - anger, sadness, disgust, any other? And e-commerce people are more interested in questions like “Does this sentence have anger?” or “Does this sentence have disgust?” To address this, we built a "One v/s Rest" classifier to predict exact emotions present in an sarcastic sentence.
• We also carried out rigorous experimentation using textual modality and proved the importance of contextual and additional information to predict emotions in sarcastic sentence.
• Tried submissions: Sarcasm_EMNLP2020_draft , Sarcasm_NAACL2021_draft

Investigating Importance of Multiple Modalities in Sarcasm Detection

• Literature Survey
• Conducted literature survey on foundations of sentiment and emotion analysis. Extracting sentiment as a quintuple from unstructured data. Studied various representations of emotions, emotion lexicons, different approaches in emotion analysis and it's applications in the real world. Explored open resources available for emoji, emoji interpretation, emoji sense similarity, and their combined applications with deep learning.
• Objective:
• To analyze the importance of multiple modalities in sarcasm detection from text. Worked on emojis considering it as one of the modalities. Emojis help us to understand the mood of the user in a better way.
• Implemented LSTM-NN and fasttext classifier as a baseline for sarcasm detection problem which had text with emojis. Conducted experiments on these classifiers by placing emojis at different positions in the text for analysing the positional importance of emojis. Analyzed the effectiveness of features from knowledge graphs like SentiWord-Net and EmojiNet. The accuracy increased and the values were close to 90%.
• After analysing the results using LIME - An Interpretability Tool, it was clear that passing additional information like emoji do help in better classifying the sentence as sarcastic or non-sarcastic.

Technology I use for Machine Learning

1. PyTorch: Machine-Learning-With-Pytorch
2. ONNX
3. ONNX Runtime

Programming

1. Python
2. C++

ML Domain & Techniques:

1. Natural Language Processing
2. Computer Vision
3. Quantization
4. Pruning
5. Node Fusion
6. Graph Optimization

Others:

1. Docker
2. Git
3. AWS

Last updated on 13/03/2025 at 08:57 p.m. IST