Vikas S

AI-Powered Design-to-Code Automation Tool (PDF to Responsive HTML)

• Product Overview: Spearheaded data science efforts for an AI product that automates the conversion of PDF-based emailers and e-detail designs into responsive HTML. This tool eliminates manual HTML development, streamlining workflows and improving productivity across design and development teams.
• Key Modules :
• Content Extraction: Utilized Azure Document Intelligence/AWS Textract to accurately extract text, images, fonts, and metadata from PDFs, ensuring seamless design-to-code translation.
• Custom Object Detection: Designed and deployed YOLO v10 models to identify key layout features such as CTAs, lines, and other design elements critical for responsive HTML.
• Generative AI Implementation: Engineered workflows using Claude, GPT and Gemini for automated HTML-to-MJML conversion for responsive HTML. Integrated advanced features like background gradient detection, CTA shape identification, and vertical text property extraction.
• Layout Structuring: Built custom algorithms to detect multi-row and multi-column layouts, providing critical insights to front-end teams for responsive HTML generation.
• Advanced Metadata Handling: Enhanced accuracy through Python-based symbol, hyperlink, and superscript extraction, enabling robust content representation.

Impact:

• Achieved 70% reduction in development time for converting varied PDFs complexities into responsive HTML.
• Improved content extraction accuracy to ensure high-quality outputs for production-ready designs.
  Adopted by 8-10 enterprise clients , each with 5+ developers , driving significant productivity gains across teams.
• Scaled to handle diverse PDF complexities, enhancing throughput and reliability of the design-to-code workflow.

ICB (Intelligent Content Brain)

• Project Overview: Developed an AI-powered tool that processes PDFs (e.g., emailers, e-details) to extract and analyze content intelligently using advanced vision and NLP techniques. This platform provides intelligent content analysis and actionable insights for end clients. Offered on a subscription-based model, it enables clients to process PDFs efficiently while ensuring high accuracy in content extraction and classification.
• Key Contributions:
• Key Message Tagging: Fine-tuned a multi-class classification model using BERT to identify and tag critical key messages, enhancing automated content analysis.
• Image/Text Semantic Modeling: Implemented transfer learning with VGG16 and fine-tuned BERT to build a robust image-text similarity model, enabling cross-modal content alignment.
• Multilingual NER Model: Utilized SpaCy and Hugging Face frameworks to extract named entities in 12+ languages, expanding the tool's global applicability for diverse clientele.
• Document Slicing: Fine-tuned a VGG16-based model to classify valid and invalid slices of documents, ensuring precise segmentation of relevant content.
• MLOps Implementation: Designed a comprehensive data science pipeline with: Data Collection: Automated workflows for ingesting and preprocessing PDF datasets. Model Retraining: Built custom pipelines leveraging AWS for iterative model updates based on feedback loops. Model Deployment: Utilized MLflow for seamless deployment and version management.

Impact:

• Achieved 70~80% accuracy across domain-specific models with iterative retraining, ensuring reliable content extraction for enterprise clients.
• Enabled seamless processing of PDFs on demand, supporting multiple clients under subscription models for PDF processing.
• Delivered a scalable MLOps dashboard that streamlined workflows for data scientists, reducing pipeline latency and enhancing operational efficiency.
  

• Binary Classification for Bad Debt Estimation: Developed a binary classification model using ensemble techniques to estimate the probability of customers being a Bad Debt. Enhanced prediction accuracy with multiple model integration. Impact: Saved ~25% revenue YoY by proactively managing financial risks.
• Time Series Modeling for Pharmaceutical Sales Prediction: Utilized the Holt-Winters model for forecasting units and sales of multiple pharmaceutical brands over 2 years. Improved accuracy by applying better seasonality and trend factors. Impact: Achieved 10-15% higher accuracy compared to the previous SAS model, improving sales predictions.
• Engineering Activities - End-to-End Automation:
  Automated legacy SSIS workflows using Python, eliminating manual steps in ETL processes. Impact: Saved 50+ hours/month , reducing 5 FTRs and improving operational efficiency.

• Worked on Predictive Modelling for internal stakeholders to estimate SLA for Priority Incidents, Resource Estimation based on SLA which enabled 20-25% better use of Resource Management.
• Led Server Infrastructure Development and Management of Production systems for ETL(informatica)
• Involved in Automation process of Admin Console Monitoring which resulted in savings of manual 15% efforts