Mushirul Aman

I designed and deployed a production-grade ChatGPT-style conversational AI application with Next.js, FastAPI, and MongoDB, with a fully automated AWS deployment pipeline as part of the AI Architect Program. I architected the end-to-end CI/CD workflow using GitHub Actions with Docker containerization, automating image builds and pushes to AWS ECR. I provisioned compute on AWS ECS Fargate (serverless containers) behind an Application Load Balancer to distribute traffic at scale, with ElastiCache (Redis) for session caching and CloudWatch and Grafana for observability. I established isolated dev/staging/prod environments using AWS CLI automation and IaC scripting. I designed the security layer using IAM roles, AWS Secrets Manager, and VPC isolation to ensure zero-trust access across all services.

I architected a production-grade multimodal RAG system capable of ingesting and semantically searching across text, PDFs, images, audio, and video using a single embedding model—Google Gemini Embedding 2—as part of the AI Architect Program. I designed the full pipeline: multimodal document ingestion; vector embedding generation via Gemini Embedding 2; storage and retrieval in the Pinecone vector database; and a conversational chat interface that enables natural-language queries over uploaded files (e.g., find the video where I discussed X, show the image with a person in black). I used Claude Code with CLAUDE.md for AI-assisted development and Ollama for local model evaluation. I assessed retrieval precision across all media types in a 100+ language-capable architecture.

I built an end-to-end LLM fine-tuning pipeline on GPU cloud infrastructure as part of the AI Architect Program, implementing parameter-efficient fine-tuning (PEFT) techniques, including LoRA, QLoRA, prefix-tuning, and direct preference optimization (DPO). I provisioned and configured cloud GPU VMs, designed dataset preparation workflows, and ran full training cycles on 7B-parameter models—reducing trainable parameters to under 0.1% of the full model. I built a FastAPI inference server for post-training model serving and conducted a comparative benchmark against RAG-based retrieval to determine the right architecture for each use case. I established a complete MLOps workflow covering environment isolation, model checkpointing, GPU cost management, and production deployment patterns.