Sarvagya Porwal AI Engineer

About Me Projects Work Blog

My Expertise

Undergraduate student at Indian Institute of Technology, Roorkee. I am intersted in Deep learning, Machine learning, Computer Vision and Generative AI

RESUME

Computer Vision

I specialize in combining computer vision and natural language processing. My focus is on image segmentation, object detection, and classification. I explore advanced techniques to improve these tasks and contribute to the field by developing innovative solutions.

Generative AI

Builiding and enhancing RAG-based pipelines, agentic architectures and prompt-engineering. Finetuning open-source LLMs for down-stream tasks.

Featured Projects

mountains

AI Agent 007: Tooling up for Success (Inter-IIT Techfest 2023)

  • Python
  • Langchain
  • OpenAI

Focused on creating autonomous tools for efficient parameter extraction and downstream function calls. Implemented a self-reflective ReAct style agent, curated dataset using given tool descriptions

Check it out
mountains

Enriched Bots-Clever Chat

  • Python
  • Django
  • Llama-Index
  • Hugging Face

Today’s bots are plain text. Our graph-based approach enriches interactions with links, pictures, and videos.

Check it out
mountains

Wheat Seeds Classification

  • Python
  • Pytorch
  • OpenCV
  • Computer Vision

The research employed RGB and hyperspectral images for classifying 96 wheat seed varieties. Pre-trained models like DenseNet-121, ResNet-50, and GoogleNet were fine-tuned for RGB data. For hyperspectral data, a DenseNet-121-inspired architecture was developed, incorporating a Sparse Band Attention Module (SBAM) to rank and select bands based on their contribution to the classifier's accuracy. This approach achieved 92% accuracy on hyperspectral images. A regression-based ensemble (using SVM) was employed to combine model predictions, enhancing robustness and achieving an overall accuracy of 98%.

Check it out

Work Experience

DeepLogic AI

Worked as a Generative AI intern.

  • Spearheaded the development of a Retrieval-Augmented Generation (RAG) pipeline for enterprise search, integrating email and document embeddings into a PostgreSQL vector store on AWS, enabling high-performance and context-aware information retrieval.
  • Designed and implemented normalized database schemas and optimized scalable CRUD operations for metadata-filtered searches, supporting seamless querying across millions of documents.
  • Engineered core components including the Retriever, Response Generator, and Re-ranker, leveraging advanced caching strategies to enhance chatbot integration, resulting in improved interaction efficiency, scalability, and user experience.

Blog

Inter-IIT Techfest 2023:

Developed "AI Agent 007," a query-aware agent capable of allocating and reviewing tool outputs.

Focused on creating autonomous tools for efficient parameter extraction and downstream function calls.

Implemented a self-reflective ReAct style agent and curated datasets using given tool descriptions.

Project link

Sentinel-2 Field Delineation:

Participated in Solafune's Sentinel Delineation competition and developed a model for field segmentation.

Fine-tuned U-Net-based models (UNet++, FPN, DeepLabV3, Mask-RCNN) and used OpenCV for polygon processing.

Built an ensemble model to stack masks predicted by base models, achieving IOU = 0.96 on patched images.

Project link