AI and Psychology Researcher | Co Founder @LifeFundies | UPenn GPN Member | MIT CSAIL Scholar | Expertise in ML & Agents

medium

devto

leetcode

codeforces

github

Devesh Kumar Gola  • Peerlist

Explored how different prompt engineering strategies impact both performance and energy consumption in LLM-based requirements classification. Used open-source models, CodeCarbon, and the PROMISE dataset to evaluate accuracy, emissions, and prompt efficiency. Built a reproducible pipeline to support sustainable NLP practices in software engineering.

CarbonPrompt - Evaluating Prompt Engineering and Emissions i

Developing a robust cross-domain segmentation pipeline by adapting models trained on Cityscapes to perform on the IDD (India Driving Dataset). Built a supervised baseline using DeepLabV3+, followed by unsupervised domain adaptation using DAFormer with techniques like feature alignment, style transfer, and adversarial training. Applied iterative pseudo-label refinement to improve performance on unstructured road scenes. Achieved significant mIoU gains, with detailed visualization and class-wise IoU analysis for both datasets.

Cross_Domain_Segmentation - Cross-Domain Semantic Segmentati

Achieved 93.84% accuracy on the FoodVision Mini dataset by developing a Vision Transformer (ViT) model, integrating EfficientNet-B2 for feature enhancement. Leveraged transfer learning from pretrained ViTs and optimised training using PyTorch, improving food classification performance significantly.

FoodViT - Food Prediction Vision Transformer using PyTorch

Developed an optimised retrieval-augmented generation (RAG) system by integrating LangChain, AstraDB, and the Wikipedia API for scalable knowledge access. Utilised the Gemma-9B-IT LLM via Groq API for high-speed inference and implemented a LangGraph-based architecture to enhance vector search, context handling, and dynamic knowledge retrieval.

RAGBotX - Multi AI RAG Chatbot using Langgraph and AstraDB

Built an intelligent resume screening system that enhances categorisation accuracy using a Random Forest classifier trained on 1.6 million features. Developed a job recommendation engine aligned with predicted categories and user profiles. Implemented a resume parser to extract key details (name, email, skills, education) and display them through a dynamic web interface for efficient evaluation.