KESHAV KUMAR

Computer Vision, OCR Projects, Generative AI, and NLP Projects.

• Built YOLOv11-based computer vision models to detect image quality issues, including blur, low-light conditions, and glare, while classifying vehicle viewpoints (front, rear, speedometer, side view), improving annotation quality by 30%, and significantly reducing manual review effort.
• Designed and deployed scalable MLOps pipelines using AWS Lambda, Amazon ECR, API Gateway, and Terraform, enabling automated infrastructure provisioning and reliable, production-grade model serving.
• Developed an OCR automation solution using AWS Textract for Philippine national ID documents, extracting and structuring key information from unstructured text, and reducing manual form-entry effort by 60%.
• Implemented intelligent post-processing and field-mapping logic to transform OCR outputs into validated, structured records, improving downstream system integration accuracy.
• Led the development of an end-to-end vehicle inspection platform, leveraging YOLOv11 for image quality assessment and viewpoint classification, improving dataset reliability by 30%, and accelerating model training workflows.
• Delivered an OCR-driven document automation system for Philippine ID processing using AWS Textract, reducing manual data-entry efforts by 60%, and streamlining customer onboarding workflows.
• Using Docker, Lambda, API Gateway, ECR, and Terraform, enable scalable and reproducible ML model deployments across environments.
• Engineered a RAG-powered Text-to-SQL assistant using LangChain, Gemini 1.5 Flash, and ChromaDB, enabling business users to query enterprise data in natural language, and reducing ad-hoc reporting turnaround time by 80%.
• Improved SQL generation accuracy to over 95% through semantic retrieval-based few-shot prompting, dynamically injecting contextually relevant historical queries using ChromaDB.
• Designed a multi-layer security framework incorporating query validation, access controls, syntax verification, and EXPLAIN-based auto-retry mechanisms, preventing destructive SQL execution, while ensuring reliable query generation.
• Built a production-grade RAG-based Text-to-SQL solution using LangChain, Gemini 1.5 Flash, and ChromaDB, achieving over 95% SQL accuracy while reducing business reporting turnaround time by 80%.

• Cleaned and enhanced ID images using custom preprocessing techniques for OCR pipeline for Indian Government IDs.
• Fine-tuned YOLOv8 to locate text and integrated Tesseract and PaddleOCR for extraction.
• Developed Flask API delivering structured JSON responses from ID images.
• Built ETL pipeline to consolidate, transform, and visualize operational data, cutting processing time by 50%.
• Created ML-driven insights dashboards using Plotly and SSRS for cross-functional teams.
• Built segmentation models from scratch by analyzing data and training models that predict high risk customer using Random Forest and XGBoost.
• Reduced marketing cost by 15% through targeted segmentation and data-driven features.

• Processed and analyzed 6 months of transactional data to identify key customer segments for transaction-based customer segmentation project.
• Created segmentation models revealing 6 distinct classes of user behavior.

• Built Flask-based API to scrape Gmail data using Google OAuth and deployed with Docker/Nginx for Gmail Data Mining API project.
• Enabled analytics to extract customer purchase patterns for personalized marketing.