How to vibe code an AI reporting tool for 100+ courses

A two-person team at the Global Surgery Foundation used AI to build a tool for their SURGhub training platform. Once it worked, they stripped every piece of AI back out. What used to take one person a week or more of full-time work every quarter, manually pulling reports course by course, now takes a couple of minutes with one click.

SURGhub sends survey data and course results to 34 partner organizations every quarter. That data has to be right every time, since it feeds decisions those organizations make about their own training. 

A tool with any chance of inventing a number, even a small one, wasn’t a risk worth taking. The most interesting part of this story is what the Global Surgery Foundation team – Sebastian Hofbauer (Programme Manager) and Cindy Moige (Project Associate) – chose to do once it worked.

This conversation happened at our LearnWorlds WOL:AI summit, and there was a lot to take from it.

When growth becomes the problem

SURGhub is the first UN-backed training platform built specifically for surgical care, hosted on LearnWorlds and launched in 2023. 

A joint initiative of UNITAR and the Global Surgery Foundation with the Royal College of Surgeons in Ireland, backed by the Johnson & Johnson Foundation, it delivers surgical, anaesthesia, gynaecology and obstetrics, and perioperative care training into resource-limited settings.

SURGhub gained a lot of recognition over the years. Its success and accelerated growth are evident by the numbers:

• More than 140 courses from 34 content providers
• More than 44,000 users, 23,000 post-course survey responses, and 21,000 certificates earned
• Average course rating of 4.6 out of 5
• Partners including the ICRC, Harvard University, the University of Oxford, and Johns Hopkins Medicine
• A Times Higher Education Award for Technological/Digital Innovation
• A QS Reimagine Education Award for Best E-Learning Project

None of this came with more people managing the reporting side of things. 

Sebastian and Cindy had to navigate to and download a post-course survey report for each course individually, every quarter, then compile and distribute the results to all 34 partner organizations. At 100 courses, that cycle took roughly a week. At o-plus, it stretched to a week and a half, since every part of the job was manual and nothing about it got more efficient with scale.

The pressure is only building. SURGhub is targeting 100,000 total users by the end of the year, roughly double where it stood at the time of this session, with recent daily sign-ups already running about 22 percent ahead of the prior month and 128 percent ahead of the same period last year.

As Sebastian says, SURGhub really has two customers: the learners taking the courses, and the 34 organizations providing them.

The providers own their courses and their data, not SURGhub, which means every report has to show each partner specifically its own learners, its own countries, its own results, rather than one generic platform-wide summary. 

Serving two audiences with genuinely different needs, on a manual process, was where the week per quarter mostly went.

Hiring an engineer to fix this wasn’t really an option. Global Surgery Foundation is a lean team running a mission-driven platform, not a software company with a backlog and a sprint board. The fix needed to come from inside the team.

What Surge Dash built to go from a week to one click

Surge Dash imports five separate exports out of LearnWorlds, centralizes them, and turns the result into a dashboard and a set of branded, one-click PDF reports for each partner.

“LearnWorlds has a lot of data. It’s incredibly valuable.”

—Sebastian Hofbauer, Programme Manager at Global Surgery Foundation

The challenge was never a shortage of data. It was that it lived scattered across the platform, in different reports, different formats, different places to click, which made it hard to centralize and turn into something usable. 

• Course, enrollment, and certificate report: the base data on who took what and finished what
• Course-to-provider map: which of the 34 partner organizations owns which course
• Course-links file: used to auto-fetch each course’s survey report without a human clicking through to it
• Learner demographics: who is taking the training
• Ambassador and referral data: tracking how SURGhub’s community ambassadors bring in new learners

One detail made the automation possible at all. 

Each post-course survey link contains a token that expires every 24 hours, but the token itself is identical across every single course. Surge Dash only needs one fresh example link each day to update and re-fetch every course’s report, rather than 140 separate ones. That’s a small technical fact with a large effect on how much manual work the tool actually eliminates.

This is the kind of problem a small team can vibe-code well: data that just needs pulling together, no judgment calls, just mechanical work a human used to do by hand.

The dashboard has picked up a second job beyond reporting to partners. 

Sebastian now pulls charts straight from it into board updates and donor conversations. A specific number lands better than a vague claim, and now he has one on hand for every conversation.

The one rule that mattered most: no AI in the finished tool

Surge Dash was built with AI, but it contains none. That distinction sat at the center of every decision the team made once the tool started working.

The data flowing through Surge Dash goes straight to 34 real partner organizations making real decisions off it. A hallucinated number in a partner report isn’t a cosmetic bug but a false claim about someone else’s training outcomes, sent under SURGhub’s name. So once a feature worked, the team’s job was to strip out anything probabilistic and lock in code that does the same calculation the same way, every single time.

If the data matters, make it deterministic, even though AI is what got you there. The tools used to write Surge Dash never touch the reports it produces once it’s running. That split, AI for building, plain code for running it, is the real takeaway here, more than any specific line of code.

How they got here: three tools, three lessons

Surge Dash wasn’t a single build, of course. It moved through three tools, and each one hit a real wall before the team moved on.

The team’s own habit mattered as much as any single tool. Staying curious and patient, and asking the AI directly about a problem, including from a screenshot of an error message, turned out to be a more reliable way forward than trying to debug it themselves first.

It’s a small habit, but it’s the difference between getting stuck at the Gemini stage and pushing through to something that actually holds up.

What it actually took

The honest number of the time needed to finalize this is roughly four full workweeks of net effort, done outside Sebastian’s regular job as a self-described passion project, not a weekend build. That’s worth sitting with before assuming vibe coding means fast and easy in equal measure.

Cindy joined the team roughly six months before the WOL:AI session and inherited the ChatGPT-and-Python version first. 

It took her a few weeks to get comfortable with it, not because she lacks technical grounding, but because that early version leaned on traditional scripting she hadn’t worked with before. The Claude-built version was a different experience entirely.

“I wouldn’t have to do that much work to get the data that I need. But previously I had to do a lot of work to get the information I needed.”

—Cindy Moige, Project Associate at Global Surgery Foundation

Prototypes come together fast. The last stretch –reliability, edge cases, something that survives a real quarterly cycle without babysitting– took disproportionately longer than the exciting first draft did. That gap is the actual cost of a tool like this, and it’s the part that’s easy to underestimate from the outside.

It’s also worth noting what this timeline doesn’t include. 

Four workweeks cover building and hardening Surge Dash itself, not the institutional knowledge Sebastian and Cindy already carried into the project. They have years of knowing exactly which numbers each of the 34 partners actually cares about, and exactly where the old manual process wasted the most time. 

A team without that context would likely spend real additional time just figuring out what to build before writing a single prompt.

Three things worth knowing before you try this yourself

Set the specifics of SURGhub’s tool aside, and three lessons here apply well beyond it.

1. Vibe coding isn’t easy despite the name

You need to deeply understand and be able to articulate your own problem before an AI tool can build the right thing. The team’s clearest advantage was that it already knew exactly what a good report needed to contain.

2. Prototypes come together fast, but the final 20%takes disproportionately longer

Treat this as their version of the 80/20 rule. Budget real time for the unglamorous stretch after the demo works, since that’s where a tool goes from impressive to actually usable every quarter.

3. It’s a niche solution, not proof that off-the-shelf software is obsolete

Surge Dash solves one specific, well-understood, spreadsheet-shaped problem for one specific team. It’s not a case for replacing purpose-built platforms broadly, and treating it that way would be reading more into this story than the story supports.

None of these three require a technical background to apply. They’re closer to project-management discipline than engineering advice, which is probably why a two-person nonprofit team could pull them off in the first place.

“Thanks to LearnWorlds for giving us this opportunity and for giving a home to SURGhub.”

—Sebastian Hofbauer, Programme Manager at Global Surgery Foundation

Use AI to build the tool, not to run it

SURGhub’s real takeaway here shows us a repeatable pattern. You can use AI to get the build done fast, then hand the actual running of it over to plain, deterministic software you can trust with real partner data. A two-person team pulled that off without a hiring budget, and the platform underneath them mattered as much as the AI on top.

If you run training for a mission-driven organization and recognize the reporting problem described here, LearnWorlds’ nonprofit training platform is built for exactly this kind of multi-partner reporting. 

Start a 30-day free trial and see what your own data looks like once it’s centralized.