Designed and implemented a 6-layer defense-in-depth guardrail architecture enforcing reliable, role-scoped information flow within a multi-agent system using a real-world supply chain dataset (DataCo Global, 180,519 orders, 909K synthetic messages). Built a LangGraph 6-node agent pipeline with LLM-powered intent classification, SQL generation, and answer synthesis via Gemini with 3-provider fallback . Agents query structured data via PostgreSQL perspective views and semantic data via ChromaDB vector store (568K documents across Sales + Logistics partitions). Achieved 100% guardrail true positive rate , 80% LLM routing accuracy , and sub-second guardrail enforcement latency (P50: 50.75ms).

Architected an enterprise-grade Agentic AI data analytics platform leveraging Large Language Models (LLMs) for autonomous data analysis and insights generation. Implemented multi-agent reasoning architecture with context-aware processing, enabling natural language data queries and automated visualization generation. Built with TensorFlow deep learning for anomaly detection using autoencoder neural networks, 3D PCA dimensionality reduction , and intelligent parsing supporting multiple file formats (CSV, Excel). Features 8+ interactive chart types with Plotly, domain-specific business insights, and a conversational AI assistant enabling real-time exploratory analysis through natural dialogue. Deployed with Docker containerization and optimized caching for production-grade performance.

Engineered a comprehensive end-to-end data engineering platform for supply chain optimization combining automated data ingestion , SQL transformations with dbt and 8 production-grade ML models for predictive analytics. Implemented XGBoost classification with SHAP explainability to predict late deliveries, Prophet time series forecasting for demand planning, and K-Means clustering for network optimization. Integrated causal inference modeling (OLS) , safety stock optimization formulas, market basket analysis and discrete-event simulation (SimPy) for scenario testing. Architected with Apache Airflow orchestration for automated daily pipelines, PostgreSQL data warehouse with staging and marts layers, and interactive Jupyter analytics notebooks . Deployed with Docker containerization enabling reproducible, production-grade performance across environments.

Engineered a comprehensive data analytics platform to address operational disruptions in automated warehouse shuttle (AKL) logistics at airport facilities. Built a modular Python-based data extraction pipeline using Design Science Research Methodology (DSRM) to parse heterogeneous semi-structured log streams and reconstruct complete, timestamped box journeys across conveyor, lift, and buffer stations. Developed ML-driven failure point localization and delay pattern detection models identifying 3,000+ potential operational delays and surfacing systemic idle wait clusters at conveyor merge zones as critical throughput bottlenecks. The platform delivered real-time operational transparency enabling early intervention workflows and data-driven preventive maintenance strategies.

Maintained GCP infrastructure using Terraform and Docker for internal e-commerce projects. Set up CI/CD pipelines to streamline deployments, covering environment promotion (dev→staging→prod) and automated validation checks. Improved deployment reliability and consistency across internal projects while maintaining infrastructure as code best practices.

Built an internal reporting tool on GCP that automated data aggregation from multi-region OMS datasets (orders, invoices, inventory, credit memos, product audit trails). Implemented role-based access controls and dashboards, reducing manual tasks for regional e-commerce teams.

Implemented an attention-augmented CNN architecture for hyperspectral super-resolution achieving 2× spatial enhancement , ~4 dB PSNR gain and >90% spectral similarity. Ranked 2nd among 20+ capstone projects demonstrating optimized trade-offs between spectral fidelity and reconstruction latency.

Executed 1,000+ Monte Carlo simulations applying multi-factor discount rate sensitivity and stochastic price curves to derive 10-year NPV distributions for a manufacturing investment case. Built an interactive Power BI executive dashboard surfacing risk bands, sensitivity sliders, and scenario comparatives to accelerate decision cycles.

Modernized a legacy core banking stack by introducing a microservices architecture using Quarkus (Java) for low-memory JVM services and ASP.NET Core for high-throughput APIs. Implemented event-driven messaging and containerized CI/CD pipelines, reducing deployment time by ~30% while improving horizontal scalability and operational observability.