- Description: A Retrieval-Augmented Generation (RAG)-based assistant designed to handle diverse compliance use cases.
- Key Achievements:
- Developed a RAG-based assistant integrating API-based models like GPT-4 and deploying fine-tuned models (SLMs) on-premise for cost-effective, customized solutions.
- Worked on fine-tuned multimodal models to support scanned documents and image-based retrieval, enabling comprehensive compliance analysis across text and visual data.
- Optimized on-premise fine-tuned models to reduce dependency on external APIs, achieving significant cost savings while maintaining high performance.
- Tech Stack: RAG, Hugging Face, LoRA, QLoRA, AutoTrain, Python, GPT-4, Prompt Engineering, Knowledge Distillation
- Key Achievements:
- Built a tool that converts natural language queries into SQL statements, enhancing productivity and data-driven decision-making.
- Pre-trained non-code-based LLMs to understand SQL contexts and fine-tuned models like LLaMA3, CodeLlama, and StarCoder using advanced techniques such as QLoRA/LoRA and PEFT for optimized inference and accuracy.
- Used a multi-agent, actor-critic framework and AutoGen/langgraph to iteratively improve query accuracy, while generating synthetic data for fine-tuning the SQL Language Model (SLM).
- Achieved over 90% accuracy on client-specific databases, significantly improving productivity and decision-making efficiency.
- 📦 Check it out: One of my fine-tuned models is live on HuggingFace—1,000+ downloads and counting!
- Description: AI-powered enhancements for CCTV systems providing live security alerts, computer vision on edge devices, privacy-preserving algorithms, and federated learning.
- Key Achievements:
- Model Development: Created models for person detection, re-identification, action classification, and depth estimation. Leveraged contrastive and SimCLR losses for unsupervised pre-training and supervised fine-tuning of person re-identification models.
- Pipeline Development: Built an end-to-end pipeline using C++ with multithreading, asynchronous coroutines, and buffers (using Folly and ZeroMQ). Designed a simple GUI with QT Creator.
- Optimization: Deployed deep learning models on edge devices (Jetson Devices) for real-time performance with high accuracy and low RAM usage. Techniques used included pruning, TensorRT, and Torch-TensorRT, along with indexing methods for re-identification similarity search optimization.
-
Generalized Method to Validate Social Distancing Using Neural Network and Median Angle Proximity Methodology
Published in ICISS 2020 International Conference, proceedings by IEEE. (Focus Area: Computer Vision) -
Design and Development of AGV
Presented at VISHWACON 2020, 3rd International Conference on Recent Trends in Engineering and Technology. (Focus Area: Robotics and Automation) -
Improved Real-Time Detection of Vehicles on South Asian Roads
Published in Engineering Science and Technology, an International Journal. (Focus Area: Computer Vision)
