
BizAI is a Streamlit-based AI agent that automates two real business workflows email replies and invoice generation using a transparent multi-step reasoning pipeline. The system is designed to move beyond single-prompt LLM usage and demonstrate a structured, agentic approach to problem-solving. The email workflow operates in three stages: first, the model analyzes the input to detect intent, tone, and urgency; second, it generates a response strategy; and third, it produces the final reply conditioned on that strategy. The invoice workflow combines LLM-based extraction with deterministic Python validation to ensure accuracy in quantities, pricing, and totals, reducing common hallucination errors. A visible agent-trace panel allows users to see each step of the reasoning process in real time. For deployment, the system uses an open-source model (Mistral-7B-Instruct-v0.2) served via a vLLM inference server running on AMD hardware through ROCm. This enables an OpenAI-compatible API interface, allowing the existing application to integrate without major code changes. Sustained throughput on a single AMD Instinct MI300X measured at ~223 tokens/sec, with the GPU at 64% utilization and 560W draw under load (rocm-smi captures included in the repo). The architecture follows a modular pattern: Streamlit frontend → OpenAI-compatible client → vLLM server → ROCm → AMD Instinct MI300X. This approach highlights the practical advantage of AMD's ecosystem: developers can run open-weight models with high performance while maintaining flexibility and avoiding vendor lock-in. The project focuses on delivering a working, end-to-end agent system that is both explainable and deployable in real-world scenarios.
10 May 2026