Hafiz Muhammad Usama Saleem

Summary

What I did

• Built an AI Voice Assistant integrating LLMs, Speech to Text, Text to Speech, Whisper, and Twilio to automate customer interactions and reduce manual support effort.
• Architected Retrieval Augmented Generation pipelines using LangChain, embeddings, vector search, and OpenSearch for document intelligence and knowledge retrieval.
• Fine-tuned and integrated open-source LLMs for domain specific applications, improving task performance while reducing inference and licensing costs.
• Developed Agentic AI and multi agent workflows using LangGraph, CrewAI, and SmolAgent for research, information extraction, reasoning, and structured report generation.
• Developed reusable, production grade FastAPI REST APIs for AI and ML services and integrated retrieval and generation capabilities into business applications.
• Built Machine Learning workflows for financial data processing, feature engineering, KPI forecasting, and automated reporting using Python, Pandas, NumPy, and Scikit learn.
• Containerized AI services with Docker and supported cloud deployment, monitoring, and production delivery using AWS, Git, GitHub Actions, and CI/CD, following MLOps best practices.

Results

• Reduced voice-to-voice conversation costs from 18-20 cents/minute to a significantly lower rate by replacing proprietary APIs with fine-tuned open-source models.
• Architected multi-agent workflows using LangGraph and CrewAI to automate complex research and reporting tasks.
• Achieved a major reduction in per-minute inference costs for AI voice interactions.
• Automated customer interactions, significantly reducing manual support effort.