Orbecta v3.3— The day I decided that better doesn't always mean more
Today I reworked a large part of Orbecta's AI layer.
Not because something was broken. Not because customers asked for it. But because stronger models have become available, and I believe a procurement platform that people base decisions on should run on the best tools available.
Yet in the end, this day wasn't about AI.
It was about judgment.
Five of Orbecta's six AI functions now run on the most capable model available. The claim analysis, Pro Scan, contract analysis, war room prep, and cluster synthesis have all been upgraded.
But one function was deliberately left untouched.
The auto-fill that extracts metadata from contracts.
Not because I forgot it. Precisely because I looked at it carefully.
Pulling a date out of a contract doesn't require a heavy model. Recognizing a contract number doesn't either. There, a more expensive model adds almost no extra quality. Only extra cost.
Maybe that was the most important decision of the day.
Not making everything better just because you can.
Asking the same question over and over again:
Does this actually add value for the user, or am I building it because I find it technically interesting?
That question kept coming back.
I decided not to build an extra nightly synthesis run. No one would read it while my customers are asleep.
I decided not to build a review function for war room preps. In practice, I'd probably never use it myself.
I decided not to further optimize a document that was already good enough.
In the past, I would have built all three.
Today I didn't.
That felt like progress.
Not because the technology got better, but because my judgment did.
There was also a moment where I had to go back into Orbecta's own history.
I discovered that two versions of the same contract analysis function existed. An old one and a new one. Both active. Both deployed.
Instead of guessing, I went back to the logs, the webhooks, and the actual calls to understand exactly which version was live.
That might sound like a small detail.
But it reminded me that Orbecta is no longer a project I know entirely by heart. It has become a product with a history, with choices, iterations, and layers built up over months.
Every now and then, you shouldn't only build new things.
Every now and then, you also have to clean up.
Today I also redesigned the rhythm of the intelligence scans around the global working day.
An early European scan.
Updates throughout the day.
And an evening scan that captures the full American trading day.
The goal is simple:
When a procurement or supply chain professional opens their laptop, the most important things should already be waiting for them.
No need to visit twenty websites first.
No need to gather news first.
No need to connect the dots first.
Just start working.
The way a morning newspaper used to work, really.
Only this time for procurement.
Finally, I reflected on something many software companies think about far too late: responsibility.
Orbecta processes contracts, supplier information, and company data. That carries a responsibility that goes beyond technology.
I store that data securely — without storage, the platform couldn't function. That's a given.
But what matters just as much is what you don't do.
I deliberately don't use that data for anything other than what the customer gave it for.
I want Orbecta to get smarter. I want the platform to generate ever better insights. But if that ever happens based on customer data, it has to happen the right way: anonymized, aggregated, and always with the customer's interest first.
Not because regulation says so.
But because trust is the most important product a platform like Orbecta ultimately sells.
What I learned as a founder today
Writing code is not the hardest part of building.
The hardest part is deciding what not to build.
Technology creates possibilities.
Judgment creates value.