BackArchive

All Projects

Research and industry work in AI, NLP, cloud architecture, and machine learning.

Agentic RAG Chatbot for University Student Services

Rebuilt a failed RAG deployment into an agentic chatbot on Azure serving 40,000+ students and staff — fixing broken document processing and retrieval, then adding conversational clarification tools, parent-child retrieval, and an offline eval pipeline to make it reliable during peak enrollment.

Claude Sonnet 4.5 (orchestration)Claude Haiku 4.5 (generation)Azure OpenAI (embeddings)+6

industry

Teams Meeting Summarization

Developed AI system using Whisper on SageMaker, integrated with LLMs for context-aware summaries.

OpenAI WhisperAmazon SageMakerAWS Bedrock Sonnet+1

industry

Private Microservice Orchestration for Organization

Orchestrated secure cross-account microservice architecture enabling seamless connectivity across organizational boundaries.

AWS API GatewayRoute 53AWS RAM+3

industry

RAG System For Organization

Designed cost-effective Retrieval-Augmented Generation system using Aurora and OpenSearch Serverless, replacing expensive Kendra to achieve 40% cost reduction while maintaining performance.

Aurora DatabaseOpenSearch ServerlessAWS Bedrock+3

industry

Smart Locker System

Orchestrated automated locker management platform integrating IoT hardware with AWS services for secure asset management.

AWS API GatewayLambdaStep Functions+3

research

UI Component Detection for IBM AS400 System

Developed solution for detecting UI components in AS400 interfaces using image processing and ML models, improving usability and integration.

OpenCVTensorFlow/PyTorchImage Processing

research

Website Testing Using LLM

Utilized LLMs to generate diverse test cases, improving coverage and reducing manual effort. Converted plain-language scenarios into Selenium scripts for automated UI testing.

LLMs (OpenAI, Claude, Gemini, LLama, Qwen)SeleniumPython

industry

Enterprise LLM Quantization for Secure On-Premise Deployment

Developed secure on-premise LLM solution using advanced quantization techniques to compress models for fast inference while maintaining enterprise security requirements.

PyTorch/TensorFlowINT8/INT4 QuantizationPost-Training Quantization+3

research

Weak Supervision

Developed labeling functions and used LLMs to enhance precision in weak supervision pipelines for transformer models with limited domain data.

TransformersOpenAI GPTClaude+4

research

Machine Reading Comprehension Using Knowledge Graph

Built knowledge graphs using fine-tuned RoBERTa models for information extraction and efficient querying methods to power search engines and recommendation systems.

Knowledge GraphsRoBERTaInformation Extraction+3