Lumber Project

Inspiration

Furniture manufacturing struggles with capacity bottlenecks and complex raw timber yield calculations. Inaccurate estimates lead to missed export deadlines and expensive demurrage penalties. Lumber was built as an AI Operational Consultant to give factory owners data-driven precision over material inventory, timber yield recovery rates, and shop-floor constraints.

What it does

Lumber is a single-scenario operational co-pilot powered by a ReAct (Reasoning + Acting) agent connected directly to a factory's Production Management System (PMS) database:

Timber Yield Analysis: Calculates raw log volume requirements based on yield recovery factors (e.g., 45%).
Bottleneck Detection: Identifies capacity constraints across CNC, Joinery & Assembly, and Finishing stations.
Pipelining & Lead Time Simulation: Models real-world manufacturing lead times, including a 12-day Kiln Drying lag.
Strategic Optimization: Evaluates trade-offs between internal production, ready stock usage, and subcontracting options.
Transparent Reasoning: Renders the agent's SQL queries and step-by-step thought process live in the UI.

How we built it

Frontend: React + Vite, Tailwind CSS, Recharts/shadcn UI for the operational dashboard and interactive chat UI.
Backend: Python + FastAPI exposing a REST API for agent execution and data retrieval.
AI Agent Engine: ReAct pattern with SQL tooling (execute_sql). Reads PMS database schema safely via read-only queries.
Database: SQLite (pms_dummy.db) schema modeling products, BOMs, raw material inventory, workstations, and work orders.

Challenges we ran into

Pipelining Lead Time Logic: Ensuring the LLM accurately calculated production schedules based on the primary bottleneck station rather than simple sequential summation.
Database Safety: Restricting the AI's SQL tools to read-only operations (SELECT/PRAGMA) to protect PMS data integrity.
Live Thought Streams: Formatting backend ReAct logs and SQL queries into structured JSON for frontend thought-process visualizers.

Accomplishments that we're proud of

Proven Case Study: Solved a 100-dining-set export scenario, shrinking production lead time from 49.5 days to 34 days to safely meet a strict 40-day deadline.
Transparent Agent UX: Built a visual Agent Thought Process panel to prove true agentic execution instead of a standard LLM text wrapper.
Decoupled Architecture: Delivered a clean, production-ready React + FastAPI microservice setup within a 5-day hackathon sprint.

What we learned

Incorporating physical manufacturing realities (kiln lag, yield loss) directly into system prompts dramatically increases domain accuracy.
ReAct loops backed by structured SQL execution outperform pure prompt engineering for operational problem-solving.
Strict API contracts enable fast, independent parallel development between AI and Web engineers.

What's next for Lumber Project

Multi-Scenario Support: Expanding beyond the furniture case study to general discrete manufacturing workflows.
Human-in-the-Loop Write-Back: Allowing factory managers to approve AI recommendations to automatically update stock allocations and issue Work Orders.
ERP & IoT Integration: Connecting the agent to live IoT kiln sensors and enterprise ERP systems (Odoo, SAP).