AI Operations Agent: RAG-Powered Retail Intelligence & Task Automation

This project was built for large-scale restaurant groups and multi-unit retail operators who manage high volumes of data across dozens or hundreds of locations. Specifically designed for Regional Managers and Operations Directors, the system serves as an enterprise-grade "Digital Consultant" that bridges the gap between fragmented POS/inventory data and daily on-the-ground execution. By transforming millions of rows of restaurant performance metrics into high-priority tasks, it provides a centralized platform for leadership to monitor KPIs, approve AI-suggested corrective actions, and ensure operational consistency across their entire portfolio. 1. What We Built We developed a production-ready Autonomous AI Operations Agent designed to bridge the gap between complex retail data analysis and daily execution. The system acts as a digital consultant for regional managers, transforming raw KPIs into actionable tasks. Analytical AI Chat: A free-form conversational interface where users can query performance data (e.g., "Show me the top 5 worst profitable stores in Istanbul for the last 3 months"). Task Management Dashboard: A structured workflow where AI-suggested actions are automatically logged for manager approval or rejection. Automated Action Logic: The agent uses an "Action Suggestion Map" to identify specific defects (like low audit scores or high food waste) and suggest precise corrective measures. Persistent Memory: Includes both short-term memory for the current chat session and long-term RAG memory to maintain context over time. 2. How We Built It (The Stack) The system was engineered for scalability and reliability using a modern, containerized stack:AI Orchestration: LangGraph was used to manage complex, multi-turn reasoning and agentic workflows. Frontend: React/Next.js 14 for a responsive, real-time user interface. Backend & Data: Node.js paired with a PostgreSQL database capable of handling 1M+ records. LLM Access: Integrated via OpenRouter to allow for flexible model selection and switching. Infrastructure: Fully Dockerized to ensure consistent deployment across environments. 3. Challenges We Faced As the system scaled from prototype to processing millions of records, we encountered several critical engineering hurdles: Response Latency: The initial monolithic prompt architecture led to response times exceeding 60 seconds, far slower than the required "ChatGPT-like" speed. Prompt Verbosity & Errors: Complex questions involving multiple variables caused the LLM to lose focus, leading to "reasoning errors" and incorrect SQL generation. Hallucination Risks: In multi-branch queries, the model occasionally fabricated data points, particularly around manager hours and performance metrics. Context Switching Bugs: The agent sometimes struggled to "let go" of a previous topic, continuing to reference an old store when the user had asked about a new city. 4. How We Solved It We re-engineered the core pipeline to transition from a single, heavy agent into a Modular Multi-Step Architecture: 75% Latency Reduction: By decomposing the main logic into smaller, task-specific nodes, we dropped processing time from 60s down to 15s. Task Decomposition & Specialized Models: We stopped using a "one-size-fits-all" model. Instead, we implemented a router that uses lighter, specialized models for SQL generation and action identification, and flagship models only for final reasoning. Granular SQL Generation: Breaking the metadata analysis into narrow sub-steps eliminated SQL hallucinations. The model now only "sees" the specific schema needed for the current sub-task, ensuring 100% accuracy. 10-Point Testing Protocol: We implemented a rigorous QA protocol that specifically verified bug fixes for context switching, task duplication, and chart coverage before final delivery.