top of page

Projects

Cheque Deposition System

  • The Cheque Clearing System (OCR) automates cheque recognition and clearing using deep learning, computer vision, and OCR. It employs a Python-based deep learning model, leveraging TensorFlow or PyTorch, to recognize handwritten text on cheques accurately. Preprocessing with OpenCV standardizes images before feeding them to the OCR model.

  • Recognized text from cheques is used to train the OCR engine, enhancing its performance over time and adapting to variations. The system stores extracted text in a database, ensuring data integrity and centralized access.

  • The entire system, including the deep learning model, is containerized with Docker for seamless deployment. Deployed on an OpenShift platform, the solution is scalable and easily managed.

  • Benefits and Impact: Improved Efficiency: Automates cheque recognition, leading to faster clearing and reduced processing times.

  • Enhanced Accuracy: Advanced deep learning models and OCR training ensure minimal errors in data extraction.

  • Seamless Integration: Easily integrates with existing financial systems for streamlined cheque clearing.

  • Cost Savings: Optimizes resources and reduces manual effort, translating to cost savings.

  • Scalability and Flexibility: Containerization allows smooth scaling based on demand.

  • The Cheque Clearing System (OCR) revolutionizes cheque processing by harnessing machine learning and computer vision. Implemented at Concord Technologies, it exemplifies the commitment to advanced technologies for operational efficiency and customer satisfaction.

Next Step (Product):

  • Object Detection in Documents: Designed and implemented object detection models to identify checkbox fields in various types of documents, significantly improving data extraction accuracy.

    Field Value Extraction: Utilized Named Entity Recognition (NER) models and pattern-based extraction techniques to extract key field values from medical documents, enhancing data processing efficiency.

    Custom NER Model Development: Built custom NER models using SpaCy library and Microsoft custom NER language service to tailor solutions for specific document types and industry needs.

    Token Classification with LayoutLM: Developed a custom multi-stack model (LayoutLM) for token classification, enabling precise extraction of structured information from documents.

    AIOps and MLOps Lifecycle Design: Architected and implemented AIOps and MLOps life cycles for internal products, ensuring seamless integration, deployment, and maintenance of AI models.

    DevOps, CloudOps, and GitOps: Focused on model deployment and code maintainability by leveraging DevOps, CloudOps, and GitOps practices, ensuring robust and scalable AI solutions.

    LLM-RAG Use Cases: Integrated Retrieval-Augmented Generation (RAG) using Large Language Models (LLMs) to enhance document processing workflows. Implemented solutions for extracting and generating contextual information from large document sets, improving the accuracy and relevance of data retrieved.

    Hardware Management: Utilized GPUs, and clusters for optimal performance and cost management.

    Performance Metrics: Applied BLEU, ROUGE, and perplexity for LLM performance evaluation.

    Use Case Development: Managed fine-tuning, model inference, and prompt engineering for high-performance LLM applications.

Breast Cancer Detection

• Requirement gathering and data transfer from cloud to local file system.
• Preprocessing/augmentation of the image dataset using python open cv - lib.
• Extracted the image features using UNET architecture in Keras framework.
• Trained and evaluated the model with the ground truth dataset.

Voicebot Product

• Requirement gathering, Micro-service architecture design using Flask, Environment configuration and Designing product workflow configuration in MongoDB.
• Audio preprocessing using sox and built a dynamic work flow based on user configuration.

Automatic Number Plate Recognition System (ANPR)

• Detect and extract the number from the license plate plus identify vehicle type and model for a parking slot.
• Data collected using web scrapping and labelled/annotated using labelme tool.
• Built and trained the CNN using the pre-trained YOLO and SSD architecture.
• Designed model pipelines and hosted that as a web service using flask.

ChatBot Product

• Requirement gathering, applied design patter to build the business flow and incorporated the ORM techniques for data handing using MongoDB.
• Text preprocessing using spaCy and BERT for word embedding.
• Introduced AI to the chatbot application by integrating that with RNN (lstm) model to predict the next user query based on the chat flow.
• Hosted the webservice as AMS ec2 instance.

Speech to text (Deep speech) / VoiceBot

• Created a recurrent neural network and extracts audio signal features as input.
• Augmentation of training data by adding real-world sound/noise to reduce the error rate of the model.
• Generate the language model with n-gram techniques to predict the text from the speech.

Enhanced Customer Profiling

• Customer Segmentation was done on the sales data to cluster the customers based on RFM by using KMeans Algorithm.
• Using Sales data, Customer Lifetime Value Prediction was done.
• Churn Prediction was used to predict the Customers who will not purchase with a plan.

Automate Ticket Classification

• Extracting the feature from text data using NLP and text pre-processing using python nltk library.
• Build a deep learning model using LSTM for the output of text data in vectorization format.

BELL.CA, Quarter Sales Plan

• Transforming the data into various stages like data cleaning, normalization, so that they can be analyzed to extract insights and improve business processes using python.
• Identification and resolution of data level issues that are impacting operation performance.
• Visualizing the data in different business approaches using a BI tool.

BELL.CA, Quarter Sales Plan

• Import data from My-SQL OLTP tables daily and load it in HIVE external tables using Sqoop.
• Develop Hive queries to replicate existing business logic and store processed data in HDFS.

Gender and Age Identification

Detect the person from the live feed. Identify the gender and age category using deep face recognition model using triplet loss function 

Business card extraction 

Detect the bossiness card from open area. Change to the straiten image like geometrical transformation. Detect the text localization in the card image. After the text identification using OCR approach for text recognition.

bottom of page