• ⁠🎓 Undergraduate at IIT BHU, pursuing Engineering Physics
• ⁠🤖 Passionate about Representation Learning, Natural Language Processing, Probabilistic ML, and Reinforcement Learning
• ⁠👯 Open to research collaborations in LLMs, Generative Models, and Quantum Machine Learning
• ⁠📫 Reach me at:
  • vansh.gupta.phy21@itbhu.ac.in
  • vansh1401876@gmail.com

Oct 2024 – Present
Exposure: LLM Fine-Tuning, PEFT, LoRA, LLaMA Adapters, Emotion Extraction, Prompt Tuning
•⁠ ⁠Researched low-resource LLM fine-tuning methods and emotion extraction from text.
•⁠ ⁠Working on synthetic interviews for depression detection using the DAIC-WOZ dataset, fine-tuned on Reddit data and integrated with emotion detection.

Aug 2023 – Apr 2024
Exposure: GANs, Graph Neural Networks, Fuzzy Logic, DeepWalk, Node2Vec, Struc2Vec
•⁠ ⁠Re-designed GraphGAN with Wasserstein Loss, improving accuracy from *84.7% → 88.57%.
•⁠ ⁠Implemented DeepWalk, Node2Vec, and Struc2Vec for node embeddings.
•⁠ ⁠Developed a Fuzzy Pre-processing Layer using a modified K-Means algorithm to boost accuracy to 88.95%.

Sep 2024 – Present
Exposure: Super Resolution, Quantization, Knowledge Distillation, LipSync, Mixed Precision Training
•⁠ ⁠Worked on inference optimization with Quantization, Knowledge Distillation, and Batch Inference.
•⁠ ⁠Applied Mixed Precision Training and Post-training Quantization on CodeFormer.
•⁠ ⁠Reduced inference time by 25%, and total forward pass time by 50%.

May 2025 – Present

• ⁠Contributing to AI for Code tooling and research problems.
• ⁠Exploring techniques for intelligent code understanding and generation.

Enriching Pre-Training Using Fuzzy Logic

• ⁠Published and accepted at an IEEE Conference
• ⁠Focused on enhancing language representation through fuzzy logic integration into the pre-training phase.

PinnedLoading

1. GPT-DecodedGPT-DecodedPublic

   An implementation of the GPT(generative pretrained transformer) model, from scratch, which produces Shakespearean text by training on the dialogues written by Shakespeare along with the GPT Encoder.

   Jupyter Notebook

2. Rupee-vs-Dollar-Time-Series-ForecastingRupee-vs-Dollar-Time-Series-ForecastingPublic

   An autoregressive forecasting implementation of a LSTM network, NBEATS architecture, ARIMA and SARIMAX regressions, and Autoformer architecture on rupee dollar exchange rates using pytorch, pytorch…

   Jupyter Notebook