Scheduler AI

Inspiration

Modern manufacturing supply chains are incredibly fragile. A single delayed shipment or a broken machine on the plant floor can halt an entire production line. Today, human schedulers react to these disruptions using legacy ERPs, spreadsheets, and endless email chains, turning them into reactive firefighters rather than proactive planners.

We were inspired by the potential of Large Language Models to do more than just generate text. We wanted to build an autonomous agent that could read real-world events (like supplier emails or floor manager chats), understand complex manufacturing dependencies, and "self-heal" a factory's production schedule instantly.

What it does

The AI Manufacturing Scheduler is an autonomous, self-healing micro-ERP system. It tracks live Work Orders, machine capacity (Work Centers), and external Purchase Orders.

Instead of waiting for manual data entry, it acts as a digital floor manager via an AI Copilot:

• Event-Driven Auto-Scheduling: If a supplier emails informing you that a component is delayed, the system intercepts the email. The Google Gemini agent parses the unstructured text, identifies the affected purchase order, cascades the delay through all dependent Work Orders, and proposes an optimized, updated schedule.
• Natural Language Copilot: A floor manager can type "The Casing Assembly Station just broke down." The AI instantly understands the mechanical dependencies, blocks the affected operations, and reroutes the schedule to prevent bottlenecks.

How we built it

We built a modern, full-stack cloud application designed for real-time responsiveness:

• AI Engine: We utilized Google Gemini (via Vertex AI) as the core reasoning engine. By providing the model with strict tool-calling capabilities (Function Calling), we gave it the agency to query our database, calculate capacity mathematically, and execute schedule mutations.
• Backend: A robust .NET 9 Web API written in C#. We used Entity Framework Core to model our complex relational manufacturing schema.
• Frontend: A dynamic Vue 3 Single Page Application styled with Tailwind CSS. We implemented Server-Sent Events (SSE) so that whenever a background webhook triggers the AI, the frontend instantly streams the AI's thoughts and live UI updates without the user needing to refresh the page.
• Database & Cloud Infrastructure: We used PostgreSQL hosted on Google Cloud SQL. The entire application is containerized using Docker and deployed as a serverless microservice on Google Cloud Run, utilizing the Cloud SQL Auth Proxy via Unix sockets for highly secure database connections.

Challenges we ran into

• Deterministic Actions from Probabilistic Models: Our biggest challenge was ensuring the AI didn't "hallucinate" manufacturing capacity. Translating a vague email into precise, deterministic SQL database updates required strict system prompts, JSON schema enforcement, and rigorous function-calling parameters.
• Real-Time Asynchronous State: Having the AI agent run in a background worker while keeping the Vue frontend synchronized was difficult. We had to build a custom NotificationChannel using Server-Sent Events to push live telemetry and agent decisions to the browser.
• Cloud Deployment & IAM: Transitioning from a local environment to a secure production cloud environment was tough. We wrestled with Google Cloud IAM permissions, custom service accounts, and ensuring that our Cloud Run container could securely connect to Cloud SQL using internal Unix sockets without exposing the database to the public internet.

Accomplishments that we're proud of

We successfully bridged the gap between a chat-based LLM and a hard, deterministic ERP system. Watching the AI autonomously parse an unstructured email, logically deduce that a specific Work Order must be delayed, and live-update the factory floor dashboard without a single human click felt like actual magic.

What we learned

We learned just how powerful LLMs can be when they are equipped with specific tools and given agency over a closed system. We also learned a tremendous amount about Cloud Run networking, Service Account IAM roles, and how to build responsive, event-driven streaming architectures in .NET 9.

What's next for AI Manufacturing Scheduler

Our next step is Multi-Agent Orchestration. We plan to introduce specialized sub-agents. For example, if the "Floor Manager Agent" detects a capacity bottleneck, it can delegate a task to a "Procurement Agent" who will autonomously draft and send negotiation emails to external suppliers to expedite raw materials.

Built With

• c#
• cloud-run
• cloud-sql
• docker
• dotnet-9
• entity-framework
• google-cloud
• google-gemini
• postgresql
• tailwind-css
• vertex-ai
• vue.js