An internship usually runs on work that is already understood. A queue of bugs, a backlog someone else wrote, the tasks that are safe to hand over. At Open Commerce, an internship is where our internal R&D happens, which works because someone a few months out of a classroom has fewer habits to defend.
Rüçhan Duman's question was one we had been carrying for a while: our project material is spread across various applications and folders, and getting an AI assistant to search all of it reliably was something we wanted and had not built. We described it to him before he had agreed to anything, so that he could weigh it properly. He is in the second year of business IT and management at Hogeschool Rotterdam; this was his first internship, and he said yes to spending all four months of it on that one question, starting from an empty folder.
Handing something that size to someone doing this for the first time only works if the program holds up its end: a written brief, a check-in every Thursday, and a team he could ask anything at any point.
Why it mattered to our clients and to us
Consulting work tends to end in a document. The client reads it, files it, and drops it into their own AI assistant when they need something out of it. From that point on, they are working from a snapshot that stopped being accurate a while ago.
We want to meet clients where they already are. If someone runs their working day through an assistant, the useful thing to hand them is a connector into the live state of their project, attached to whichever AI solution they already use, instead of a report that starts aging on delivery. That is where we want this to land, and building one for ourselves was the way to learn what it costs to build one for anybody else.
What Rüçhan built
Register a Google Drive folder, run the ingest, and the files are copied into a vector database, purpose-built for AI search, while the originals stay exactly where they are. That refresh runs every day, so the answers come from what is in the folders now rather than from a copy made at some point in the past. Ask a question and the answer arrives with the documents behind it listed underneath. When the answer sits nowhere in the knowledge base, it says so and offers to look elsewhere, instead of filling the gap with a hallucination. He also built the same connection for Jira, so the tickets sit in the search alongside the documents.
It works in all three places we use Claude: Rüçhan built it against the desktop app first, running on his own machine, then published it so that the browser version needed nothing more than a URL. The terminal took a round of its own, with commands to run, an install, and a separate authentication step before anything would connect.
The first version answered well, as long as every question opened with an instruction to use the knowledge base. Rüçhan took the time to rewrite the instructions that tell Claude what the tool is for and when to reach for it, so that step was solved.
How the work ran
From the first week he worked with real project material, testing the tool with the kind of questions we ask on live projects rather than with sample files. At the Thursday check-ins, he proposed features he had found himself needing, like a way to remove a document from the knowledge base once it had been added. Some went into the build as he suggested them, and for others the team came back with a different approach. In between, he posted a daily update in Slack: what was finished, what was next, and where he was stuck, with a name attached to each blocker so the right person could pick it up.
"When you have a question, the person who knows about it always replies. And they take a look with you. They say, if you want, we can schedule a meeting, so I can help you with every step you have difficulty with."
Rüçhan Duman, intern at Open Commerce, April to August 2026
Those calls turned out to be his favorite part of the internship, including the early ones he arrived at still half asleep, where the words would not come until a colleague suggested starting the mornings with a coffee.
Where it stands now
At the end of the build, Rüçhan presented the tool to the team, his first time doing that for a company. He is giving the same presentation at university, so he and Sander are rehearsing it together.
He wrote the documentation for connecting to the tool and using it, along with a troubleshooting guide, and handed the whole thing over to Sander before he left. The team has been working with it since; it saves a good deal of hunting around, and it turned out to be a solid enough foundation that we are still building on it.
What Rüçhan Duman built is also a proof of concept for our clients: the same thing pointed at their project instead of ours, so their own AI-powered assistant answers from what is true today.
