Shashwat SHARMA

• •Conducting research on machine learning architectures applied to high-frequency limit order book data
• •Developed and evaluated transformer architectures, studying tokenization, positional encoding, loss functions, and noise robustness
• •Leveraged full predictive distributions to assess gains over classical baselines and internal state-of-the-art models

• •Developed a price prediction model for microelectronic components, devising a novel estimation strategy that outperformed state-of-the-art models
• •Benchmarked ML algorithms and optimized XGBoost for accuracy and inference speed
• •Implemented a TabNet-based architecture with pretraining and finetuning, and constructed a knowledge graph with GNNs for enhanced regression

• •Created Jupyter Notebooks improving curriculum with 20% boost in feedback scores
• •Built automated scoring system for Python lab tests
• •Collaborated with industry experts to design real-world masterclasses

• •Exploring Machine Unlearning on Fine-Tuned LLMs
• •Implementing selective forgetting techniques while maintaining model performance
• •Experimenting with PEFT methods and various unlearning approaches

• •Designed live transcription and summarization system using transformer models
• •Reduced word error rate by 8% through fine-tuning multilingual speech recognition
• •Achieved SOTA results in medical image classification with 5-10% accuracy improvement
• •Integrated pre-trained models like VGG, ResNet, and BERT

Built MagnOSS, an AI-powered interactive chess coach integrating OpenAI models, a custom RAG pipeline, and Stockfish — developed in just 24 hours.

Developed an MCP server enabling natural language deployment of GitHub repos, VM orchestration, and Docker container management using the Mistral Medium model.

Created a JavaScript game under the theme "You don't control the character", with Mistral models powering indirect player influence and humorous AI-driven interactions.

FeaTrim-ViT: Vision Transformer Trimming with One Shot Neural Architecture Search

Sharma, S., Kaushik, B. (2024) - Fifth International Conference on Computing, Communications, and Cyber-Security