A powerful Document Question-Answering system built using Retrieval-Augmented Generation (RAG) architecture, FAISS Vector Search, and a FastAPI backend. It supports both Bengali and English for querying documents and provides context-aware answers. The system is also integrated with Streamlit for an interactive, user-friendly interface.

 Firstly, generate the text chunks from PDF documents using the Langchain text splitter. the FAISS vector was generated using the text chunks. For the embedding of the text, l3cube-pune/l3cube-pune/bengali-sentence-similarity-sbert model was used, and the llama-3.3-70b-versatile LLM model was used using the Groq API Key. For the demonstration of the project using a web application, FastAPI was used to generate the API endpoints. User can ask to upload a PDF from the web application and ask different types of questions using English or Bangla, and the API will give the response based on the information that belongs to or is not in the document 

Tech Stack: LLM, RAG, Llama 3.2, FastAPI, FAISS, Docker

The research-based project was a FastAPI-based application for real-time speech recognition using WebSocket. It processes audio data streamed from the frontend, transcribes it using OpenAI's Whisper model, and sends the transcription back to the client in real time. 

The audio was converted to a base64 format from the frontend, and it was decoded as WebM format on the server side.  HS-256 Encryption and Decryption were used to encode the client secret id and process the unique ID. A noise reducer was used to reduce the noise from the environment, and voice activity detection was used to filter the audio. 

Tech Stack: Real-time speech recognition, Whisper, Noise reducer, voice activity detection, Docker

Bangla punctuation correction is one of the most challenging tasks in Bangla language processing. In this research, Bangla sentence punctuation was corrected using two different approaches. The first one was a transformer-based model approach, where a Bangla custom dataset was trained with different types of books, news, and articles. An adversarial attack was applied to increase the dataset size. Then, I applied a BERT-based transformer model to it, and it achieved a good result. 

Finally, the LLAMA 3.2 model was used to check the performance for correcting the Bangla sentence. LLAMA 3.2 performed well for a rare type of non-punctuated Bangla sentence. The application was developed using FastAPI and deployed using Docker. 

Tech Stack: Sentence restoration, transformers, LLAMA 3.2, FastAPI, Docker

This research was a real-time wrong-side vehicle license plate detection using YOLOv10, EasyOCR, and a dynamic line drawing algorithm. Firstly, a YOLOv10 model was trained with different categories of vehicle datasets, such as bus, truck, motorcycle, and car. then another YOLOv10 model was used to extract the license plate region from that vehicle. the region of the extracted portion was passed to the EasyOCR, and finally, the OCR gives the license plate number. 

For the definition of wrong lane, a new algorithm was proposed for drawing the wrong lane of a road. if a vehicle is passed to the lane that only detects a violation of the road rule, as automatically detects its license plate and saves this information into a MongoDB database. The application was developed in FastAPI. 

Tech Stack: Real-time license plate detection, YOLOv10, EasyOCR, MongoDB, FastAPI

Bangladesh is one of the top ten rice-producing and consuming countries in the world. China, India, and Indonesia are the top three rice-producing country in the world. To ensure the healthy and proper growth of rice plants, it is essential to detect any disease in time and prior to applying the required treatment to the affected plants. Identifying diseases from the image of the plant is one of the interesting research areas in computer vision and the agricultural field. There are some popular rice plant diseases, namely Bacterial leaf blight, Brown spot, and Leaf smut. The process of plant disease detection is divided into two parts called image processing and machine learning. Various steps of image processing include image background removal, noise removal, image resizing, image segmentation, image features extraction, while machine learning includes feature selection and classification.

 Tech Stack: Disease detection, image processing, K-Means, Segmentation, OpenCV