Panenco

Panenco We’re a boutique product studio in Belgium.

We work shoulder-to-shoulder with our clients on next-gen B2B SaaS products and enterprise AI solutions, in a spirit of joint entrepreneurship.

We're moving towards an end-to-end software factory.Like most teams, we started with AI as an assistant. Engineers wrote...
02/11/2026

We're moving towards an end-to-end software factory.

Like most teams, we started with AI as an assistant. Engineers wrote code faster, but review, testing and support didn't speed up with them, so that's where work started to wait.

Going from AI-assisted to AI-operated means agents take on the steps around the code as well:
→ Knowledge, so they follow our company rules, coding standards and architecture decisions
→ Skills, so they can draft PRDs, build prototypes in the real stack, slice tickets and write tests
→ Gates on every pull request: a review agent, end-to-end tests, and security and accessibility checks
→ Automations that run on their own, like dependency upgrades, fix PRs for production errors and release notes

Engineering is where we started, but the same shift reaches further, from how we define functionality with clients to how we support customers after release.

Our engineers still decide what gets built, review anything risky and do every merge themselves. We also help other companies make the same move with their own teams.

Curious what that could look like for your team? Happy to talk you through it.

Whenever we put an AI assistant on top of a company's own documents, the model turns out to be the easy part. The files ...
26/10/2026

Whenever we put an AI assistant on top of a company's own documents, the model turns out to be the easy part. The files we get are scanned, laid out in two columns and full of tables, and the clause that answers the question tends to sit around page 40.

If you split that text every N characters, which is what most tutorials suggest, you get tables cut down the middle and fragments that no longer say what they are about. We built it that way first and spent a while wondering why retrieval kept returning passages that looked relevant and explained nothing.

Now we let the document set the boundaries, using for example Google Document AI's Layout Parser, which reads the real layout of titles, headings, paragraphs and tables and cuts along those.

Every chunk keeps the headings it sat under, so a line about payment within 30 days still knows it belongs to the commercial terms when it comes back on its own, and it keeps its page number, so a citation lands on the right page rather than the right neighbourhood.

Most of the answer quality was decided before the model ever ran.

Meet the team: Thomas Callens, one of our software engineers in Leuven.Thomas didn't take the usual route into software....
19/10/2026

Meet the team: Thomas Callens, one of our software engineers in Leuven.

Thomas didn't take the usual route into software. His background is in mechanical engineering, with a PhD in robotics spent building control algorithms for exoskeletons, before the software side won him over.

We asked him how he ended up here, what his work looks like, and what he gets out of it.

Good to have you building with us, Thomas.

"Buy unless you have no other option" has been the enterprise software rule for about a decade. It was good advice. Buil...
12/10/2026

"Buy unless you have no other option" has been the enterprise software rule for about a decade. It was good advice. Building meant a long project, a dedicated team and a system to maintain for years. A license moved that risk to a vendor.It also came with a trade most companies accepted without much debate: working the way the vendor works.

That rule deserves a second look, because the cost of building has dropped fast. AI-assisted engineering takes care of a large share of the implementation effort, and a small team now delivers in a quarter what used to take a programme a year. More on that in a next post.

What you get for building has not changed:
The terminology is yours, and screens follow the order in which the work actually happens. The exceptions, workflows and integrations that make the process yours are part of the system, instead of workarounds around it.The roadmap is yours. When something needs to change, the discussion is about what to build next, not whether a vendor will prioritize it. No dependency on a vendor's pricing, direction or future.

Buying still wins for generic processes, think CRM, ERP or payroll, where a mature product has been refined for years and the certifications and updates come included. The comparison changes when the process is specific to how you operate.

argenx ran that comparison recently, against a credible vendor with clear requirements. Building came out ahead on cost, and further ahead on fit and control. The result is a system that encodes how they run a core part of their business, owned rather than rented.

We wrote up what we learned in the blog on our website. https://www.panenco.com/updates/evaluating-the-buy-vs-build-decision-in-the-age-of-ai

Last Wednesday we hosted the second evening of our "Building in the age of AI series" in Leuven, centred on the role of ...
05/10/2026

Last Wednesday we hosted the second evening of our "Building in the age of AI series" in Leuven, centred on the role of product management.

One theme kept coming back: product and engineering are moving closer together. Engineers make more product calls as they build, and with multiple agents and engineers shipping in parallel, a working prototype can appear in an afternoon. That makes the product manager more important. Someone has to hold the bar for quality, decide what counts as done and keep the product coherent while everything moves at once. The tools will happily build anything. They won't tell you whether it should exist or whether it's good enough to ship.

We didn't arrive with answers. The evening ran as an open discussion, and the room did most of the work. People shared how their teams are shifting, challenged each other's assumptions and asked the questions they're still figuring out. Hearing those different perspectives side by side was the most valuable part of the night.

Thanks to everyone who joined and spoke so openly. Our next evening is on Thursday 21 January, with a more technical angle on AI deployment. If you'd like to join the conversation, feel free to drop us a message.

We recently built an assistant to automate customer success for one of our clients. We implemented the following archite...
29/09/2026

We recently built an assistant to automate customer success for one of our clients. We implemented the following architectural pattern for the retrieval.

Our retrieval assistant should decide whether it can answer before it actually starts writing the response. Streamed text cannot be retracted, so a check that runs after generation can only annotate an answer the reader already has. Ours runs the check first, before the LLM is called at all.

Every question goes through two searches, one on meaning and one on keywords, merged into a single ranking. How close the best match sits decides whether we answer. If it falls short, the question never reaches the model. It goes to a human queue with the closest articles and the reason it stopped attached. If it clears, the sources appear first and the answer streams in with a link to each one.

The cut-off was measured on a labelled set of real questions rather than set by feel, and it leans towards escalating over answering wrongly.

Meet Milo Wera, Product Manager at Panenco.Milo is driving QIT, a platform that helps mental health professionals manage...
22/09/2026

Meet Milo Wera, Product Manager at Panenco.

Milo is driving QIT, a platform that helps mental health professionals manage their practice. Think client records, session tracking, and all the administrative overhead that gets in the way of actual care. The goal: less time on paperwork, more time with patients.

He's at his best in the definition phase, where the real product thinking happens.

Glad to have you, Milo.

For the support layer of one of our platforms, the knowledge base mostly writes itself now.Once a week a job reads the s...
15/09/2026

For the support layer of one of our platforms, the knowledge base mostly writes itself now.

Once a week a job reads the support conversations, groups the ones that turn out to be the same question in different words, and drafts an article for whatever keeps coming back, with the original conversation attached for whoever reviews it.

What we like is that the articles come from what people asked rather than from us guessing which topics need explaining, so the gaps announce themselves: a question that keeps escalating is a missing article. The drafts still need work, some of them a lot of it, so an admin reads every one before it goes live. The support queue is our content backlog now.

We block lunch every two weeks to sit down as a team and go through what people learned so the whole team can benefit, w...
07/09/2026

We block lunch every two weeks to sit down as a team and go through what people learned so the whole team can benefit, we call this Engineering Labs.

This week, one of our engineers walked us through some of the new tooling he'd been using to multitask across repos, running several coding agents in parallel with Herdr ( https://github.com/herdrdev/herdr ) and firstmate ( https://github.com/kunchenguid/firstmate ) and letting cubic do a first review pass on the pull requests. He was upfront about the parts that still annoy him, and about where the tools genuinely help him today.

That is what these lunches are for. On your own, you stick to what you know until something forces you to look further. By having regular sessions on best practices, you find out what a colleague has been doing for weeks.

On September 30th we host the second roundtable of our "Building in the age of AI" series. For those who are eager to show what you've been working on or learn from others in the field, definitely welcome to join.

Broken access control sits at number one on the OWASP Top 10. So authorization is a good place to start if you care abou...
04/09/2026

Broken access control sits at number one on the OWASP Top 10. So authorization is a good place to start if you care about security.

Yet, in a lot of codebases it's the least designed part. A role check here, an if-statement there, added wherever someone happened to need one. It works until it doesn't, and by then nobody can say with confidence who can reach what.

At Panenco we design authorization as a system before writing a single check. Our default is additive roles. The idea is simple. Every resource declares the roles that grant access to it. A user holds a set of roles, and those roles add up. Anyone who holds the roles a resource requires can reach it, no more and no less.

What that results in:
- Access rules live with the resource, not scattered across controllers
- Granting access means assigning a role, not editing code
- A user's permissions are just the sum of their roles, easy to audit
- A new resource picks up the right users the moment it declares its roles

Flexible, centralized, and easy to reason about. Authorization is a design decision, not an afterthought.

Adresse

Hyphen One Kolonel Begaultlaan 1A/51
Leuven
3012

Téléphone

+32497863605

Notifications

Soyez le premier à savoir et laissez-nous vous envoyer un courriel lorsque Panenco publie des nouvelles et des promotions. Votre adresse e-mail ne sera pas utilisée à d'autres fins, et vous pouvez vous désabonner à tout moment.

Contacter L'entreprise

Envoyer un message à Panenco:

Raccourcis

Partager

Services Informatiques Et électroniques à proximité