Balakumaran Sugumar

Description: Architected and implemented a serverless tokenization system designed to protect sensitive financial data during digital transactions. The platform replaces sensitive cardholder data with unique tokens in real time, ensuring privacy, performance, and fault tolerance at scale. The system leverages AWS Lambda, API Gateway, and DynamoDB to deliver sub-100 ms response times across high-volume workloads.

The design supports multi-tenant isolation for enterprise clients, with centralized key management through AWS KMS and automated key rotation. Advanced observability and audit logging were implemented via CloudWatch and EventBridge, enabling full traceability across API transactions.

Key Highlights:

• Achieved serverless scaling with zero-downtime token generation.
• Reduced API latency by 45% through parallelized token lookups.
• Integrated caching layer for high-frequency merchant transactions.
• Optimized cost and resilience through event-driven architecture.

Technologies: AWS Lambda, API Gateway, DynamoDB, KMS, EventBridge, CloudWatch, Spring Boot (microservices)

GitHub: View on GitHub

Demo: Live Demo (coming soon)

Description: Designed and deployed an AI-driven transactional fraud detection system that monitors large-scale financial transactions in real time. The system leverages Apache Kafka for streaming ingestion, Apache Spark for distributed data processing, and XGBoost for model-based anomaly detection, achieving 97% accuracy while maintaining sub-150 ms response times.

The architecture supports continuous learning through Airflow and MLflow-based retraining pipelines, ensuring that models adapt dynamically to new fraud patterns. Results are integrated into a Flask API service for real-time scoring, alerting, and reporting.

Key Highlights:

• Built an end-to-end real-time fraud detection pipeline using Kafka, Spark, and XGBoost.
• Achieved 97% detection accuracy with minimal false positives.
• Deployed scalable inference services via containerized APIs.
• Integrated self-learning loops with MLflow for automated model versioning.

Technologies: Apache Kafka, Apache Spark, XGBoost, Airflow, MLflow, AWS S3, Python (Flask API)

GitHub: View on GitHub

Demo: Live Demo (coming soon)

Description: Developed an agentic AI system that autonomously assists customer queries, integrates seamlessly with CRM platforms, and optimizes support workflows using large language models (LLMs). The system is capable of understanding intent, context, and sentiment — dynamically adapting its tone and actions for human-like interactions.

Designed to act as a virtual service representative, it can perform complex tasks such as case creation, escalation routing, and post-interaction analysis. Built with a modular architecture, the agent connects APIs across Salesforce, Zendesk, and ServiceNow, reducing average handling time (AHT) by 40%.

Key Highlights:

• Implemented a multi-agent orchestration layer using LangChain for task planning and delegation.
• Integrated LLM-driven dialogue engine capable of context retention across customer sessions.
• Reduced manual resolution time by 40% through intelligent automation and response synthesis.
• Seamless integration with enterprise CRM and ticketing workflows for live deployments.

Technologies: Python, LangChain, OpenAI GPT APIs, FastAPI, Redis, Salesforce API, ServiceNow API

GitHub: View on GitHub

Demo: Live Demo (coming soon)

Abstract: This paper presents a scalable vault-based tokenization architecture that delivers sub-70 ms latency for payment token generation and retrieval. The design minimizes PCI DSS scope by 75% through a separation of sensitive data domains, while maintaining end-to-end encryption and auditable security compliance across payment channels.

The architecture combines AWS KMS for key management, DynamoDB for fault-tolerant vault storage, and Lambda for on-demand scaling. Through extensive benchmarking, the system achieves high availability, regulatory traceability, and PCI Level 1 security assurance.

Key Highlights:

• Reduced PCI DSS compliance scope by 75% through token-domain isolation.
• Achieved sub-70 ms latency for token generation and detokenization workflows.
• Implemented fault-tolerant vault clusters with automated key rotation.
• Integrated auditable access logs for compliance and forensic analysis.

Publication: Presented at IEEE International Conference on Computing, Communication, and Automation (ICCCA), 2025

GitHub (Prototype): View on GitHub

DOI: Coming Soon (IEEE Xplore)

Abstract: This paper presents an explainable AI framework that integrates LSTM neural networks with SHAP-based interpretability to detect and explain anomalous card transactions in real time. The proposed framework bridges the gap between high-performance deep learning models and the need for transparency in financial fraud detection systems.

The model captures temporal dependencies in transaction patterns using LSTM layers and applies SHAP (SHapley Additive exPlanations) to identify key behavioral features influencing each prediction. Experimental results on large-scale financial datasets show an accuracy of 96.8% and reduced false positives, while providing interpretable insights for auditors and compliance teams.

Key Highlights:

• Combines LSTM-based sequence modeling with SHAP for transaction-level explainability.
• Provides real-time interpretability for AI-based fraud detection systems.
• Achieved 96.8% accuracy and improved trust through transparent decisioning.
• Enables regulatory-aligned AI adoption in financial fraud prevention.

Publication: International Journal of Computer Science and Security (IJCSS), 2025

GitHub (Implementation): View on GitHub

DOI: Coming Soon (IJCSS Online)

Abstract: This paper presents a serverless, AI-powered fraud detection framework designed for real-time identification of transactional anomalies in financial systems. The proposed architecture leverages event-driven computing and machine learning models to deliver scalable, low-latency fraud detection while maintaining strong operational efficiency across high-volume data streams.

The system integrates AWS Lambda, Kinesis Data Streams, and SageMaker for ingestion, inference, and continuous model updates. It achieves millisecond-level decisioning latency while dynamically adapting to transaction spikes. An ensemble of supervised and unsupervised AI models enables adaptive detection of evolving fraud patterns. This architecture demonstrates how serverless cloud design can deliver enterprise-grade AI performance with minimal operational overhead.

Key Highlights:

• Developed a real-time serverless fraud detection system for scalable transaction monitoring.
• Achieved sub-100 ms latency using event-driven compute with dynamic scaling.
• Implemented self-learning fraud models with feedback-driven retraining loops.
• Optimized cost and observability using CloudWatch and EventBridge integrations.

Publication: Presented at IEEE Boston Conference on Cloud Computing and Intelligent Systems (2025)

GitHub (Implementation): View on GitHub

DOI: To be published via IEEE Xplore