Summer Projects ran for twelve weeks from November to February. Around thirty students, five teams, and a showcase night at the end where every team demoed to a judging panel.
The rule we set at the start was that each team had to present something that ran. Not a slide deck describing what they would have built with more time — a working system, live, in front of people.
What got built
- Knowledge agent — retrieval-augmented question answering over the club's own document corpus, so the answer comes from our material rather than from the model's guess.
- Marketing agent — drafts campaign content in the club's voice and queues it for a human to approve.
- Secretary agent — lives in our Discord and handles the admin loop: repeated questions, reminders, follow-ups.
- Logistics agent — plans and tracks the moving parts behind running an event, from rooms to run sheets.
- Data analyst agent — takes a plain-English question about a dataset and returns the analysis and the chart that answers it.
Each has its own write-up if you want the detail.
What we'd change
Three things came out of the retrospective.
Scope earlier. The teams that struggled weren't short on ability — they were carrying a scope decided in week one and never revisited. The teams that shipped cut something in week six.
Evaluation isn't optional. Several teams discovered late that they had no way to tell whether a change had made their agent better. "It seems better" is not a result, and building the harness first would have cost a week and saved three.
Demos need rehearsing. Obvious in hindsight. A system that works perfectly on the developer's laptop and fails in the room is, on the night, a system that failed.
The pattern underneath
What strikes us looking back is that almost none of the difficulty was in the modelling. It was in the plumbing: giving the agent real tools instead of letting it guess, deciding when it should stay quiet, keeping a human in the loop where a human belonged.
That's a reasonable summary of what building with AI is actually like right now, and it's the thing you can only learn by shipping something.