Arish Amin

Summary

What I did

• Built a real-time fraud detection model to minimize fraudulent payments by tracing user patterns and scoring actions.
• Developed an automated pipeline using Azure Functions to hash data, attach metadata, and trigger ML training jobs upon receiving new data in Azure Blob storage.
• Integrated the fraud detection API endpoint into the company codebase to protect sensitive actions.
• Built API endpoints to handle user queries and connect the retrieval pipeline with the LLM for the RAG chatbot.
• Created T-SQL scripts to auto-migrate data to Azure Blob for the fraud detection project.
• Defined expected JSON schemas and validated LLM responses with Pydantic to ensure reliable downstream integrations for APIs and business workflows.
• Implemented error handling for malformed or incomplete LLM outputs, including retry logic when responses failed validation.
• Handled multi-step requests by maintaining conversation history and passing relevant context into subsequent LLM calls.
• Orchestrated service calls using FastAPI endpoints while tracking workflow state, completed steps, and API results separately from the conversation.
• Implemented conditional logic to handle failures and determine next actions to prevent repeated steps and ensure context preservation.
• Optimized LLM applications at the orchestration and inference levels.
• Improved prompt efficiency and reduced unnecessary context in LLM calls.
• Validated outputs to avoid repeated requests and downstream failures.

Results

• Reduced customer-query response time by 60%
• Achieved 94%+ accuracy on a deep-learning fraud detection model on Microsoft Azure