MakerLAB AI

The problem

Academic makerspaces run on operational knowledge (manuals, setup procedures, safety rules, inventory, and repair history) that sits in shared drives, equipment websites, and staff memory. At the Cornell Tech MakerLAB, students arrive at every skill level and staff are not always on the floor, so that knowledge reaches them unevenly. The same three needs keep coming up: operate a machine, fix it when it breaks mid-job, and build a project across several machines.

What I built

MakerLAB AI is the assistant inside MakerLAB Tools, a web platform in use at the lab. It keeps tools, individual units, locations, resources, maintenance tickets, and projects in a SQL database with file storage, with an optional one-way sync to Notion where staff already work.

• Grounded answers. Manuals and SOPs are archived, OCR'd when scanned, split into passages, and indexed for hybrid keyword-and-vector search with reranking. The assistant cites the page it used and says so when nothing supports an answer, sending students to staff for safety and sign-offs.
• Proactive maintenance. When a student describes a problem, the assistant drafts and files a ticket against the right unit and walks them through the fix when the manual covers it.
• Research agents for inventory. Staff add equipment from a photo, a note, or a pasted list. Durable, retried background workflows research each tool (manuals, specs, product image) in under a minute, and nothing is published until a person approves it.
• Starting from the machine. Printed QR labels open each machine's page, manuals, and assistant. A photo of a label or of the machine itself tells the assistant which tool a student is standing at.
• Kiosk and MCP. A full-screen kiosk shows machine status and open tickets at the lab entrance. The catalog, manuals, and everyday actions are also served over the Model Context Protocol, so visitors can use the lab from Claude or ChatGPT.
• Multilingual. The assistant answers in each student's preferred language.

Important product decisions

• Person in the loop. Every change the AI proposes waits on a confirmation card, and after reading outside text the server, not the prompt, refuses changes to people or anything irreversible.
• The lab's SOP is the reference. Each lab adds its own SOPs and notes alongside the manufacturer's manuals.
• Spec-first. Each feature starts as a written design and ships with automated tests run on every change.

Results

In a spot check of ten student questions, eight got a complete, correct answer (four citing a manual page), two were partial where the records lacked data, and none was unsafe. Costs are in cents: about 0.25¢ per chat turn that searches a manual, 2–5¢ to research a new tool, and 1.3¢ to index 20 manuals. The project was accepted as a demo at the International Symposium on Academic Makerspaces (ISAM) 2026, with Niti Parikh (Director, Learning Spaces & MakerLABs) and Miguel Ramirez Peraza (MakerLAB intern).

What's next

The insights page logs what is asked and what goes unanswered; the next step is measuring ticket volume, machine downtime, and after-hours questions, then AI planning for digital fabrication across the lab's machines and a framework other labs can adopt.