Panenco

Name: Panenco
Address: Hyphen One Kolonel Begaultlaan 1A/51, Leuven, 3012, BE
Telephone: +32497863605

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.

04/06/2026

Every team building serious software needs someone who can hold the technical vision together while the work gets done.
At Panenco, that's Johan Haest.

As one of our Engineering Managers, Johan sits at the intersection of technical architecture, developer workflows, and business alignment. He cares deeply about building things the right way, and about the small process changes that quietly make a big difference in speed and quality.

Glad to have you on the team, Johan.

02/06/2026

Q7 Leader helps managers evaluate and reward people fairly. The challenge: defining and benchmarking roles was slow, manual, and dependent on analysts translating job information through spreadsheets and lengthy intake forms.

So we built an AI layer that does it for them.

Users describe a role via voice or chat and the AI structures it into a complete Role Description Profile, instantly matched against external compensation databases. Flexible mapping handles roles where no direct market equivalent exists.

A few things that made it work:

- Starting with rapid prototypes let us validate conversational RDP creation against Q7's real compensation logic before scaling the architecture
- We spent serious time upfront mapping Q7's methodology before touching any model code. You can't teach a system to profile roles if you don't understand the framework yourself
- The goal was never to replace compensation analysts. When matching runs automatically, analysts get their time back to focus on what creates real value: guiding customers through adoption, advising on edge cases, and deepening the relationship rather than wrestling with data entry

27/05/2026

When building AI solutions with our clients, observability is the key to success. We always keep prompts versioned, store domain expertise and playbooks, and track output quality consistently.

This knowledge lives in an evals tool, Langfuse or similar, co-managed between engineering and the people who actually know the domain. Strategy and customer-facing teams can update language and guidance directly, without touching the codebase. No engineering bottleneck every time the business logic shifts.

This turns client know-how into an operational asset. Updates roll out systematically, output shifts have a clear audit trail, and the system gets smarter in the same rhythm as the client's understanding deepens.

12/05/2026

For one of our clients, we automated the processing of incoming purchase orders. Reading PDFs and keying data by hand is now handled by an AI-driven dual-extraction pipeline that runs reliably across a wide variety of order layouts.

A Python preprocessor extracts field values using deterministic rules. Azure Content Understanding runs its own extraction on the same document. Two independent reads, each with different strengths.

Both results are passed into an LLM refinement step alongside the original document. The model cross-references candidates, resolves conflicts, and produces a single set of final values: customer name, delivery address, line items.

As a final safeguard, an LLM judge reviews the output against the original PDF and flags anything that looks off before the data moves downstream. A built-in quality gate. Each layer focused on what it does best: structured extraction where rules work, AI where they don't, and an independent check to catch what slips through.

07/05/2026

IFS (International Food Services) delivers professional catering and food services in some of the world's most demanding operational environments. Their challenge is training cooks efficiently in order to maintain consistent food quality standards without an expert watching every workstation, every session. Manual observation doesn't scale. So we built an AI layer that does it for them.

Video and multi-source sensor data are combined to map every kitchen action to specific culinary techniques. The result: exhausting real-time monitoring replaced by structured, reviewable quality insights. A few things that made it work:

- Starting with a single-session MVP let us validate AI accuracy against real culinary expertise before expanding scope
- We spent serious time upfront understanding the training ecosystem before touching any model code. You can't teach a system to recognise good technique if you don't understand it yourself
- The goal was never to replace trainers. Freeing them from constant observation so they can focus on coaching is where the value sits

Raw video in, actionable insight out.
A big thanks to Videofy and Mojuice for the great collaboration.

30/04/2026

We're moving to monorepos at Panenco. Here's why:

For years we kept it simple: one repo for the frontend, one for the backend. It worked well enough. Until it didn't anymore. As AI coding tools become part of every developer's workflow, context is everything. When your coding assistant can see both your API and your frontend in a single workspace, it understands your entire system. It catches contract mismatches between backend and frontend before you do. That alone is worth the switch. But the benefits go well beyond AI:

- One PR for a fullstack feature instead of two
- No more syncing branches across repos
- Shared tooling, shared CI, shared standards
- Developers work across the stack without context-switching
- Whole team expands skillset and becomes more full-stack oriented

Monorepos paired with AI-powered development are a genuine multiplier across our projects!

22/04/2026

For one of our recent projects, we needed to classify images reliably without the cost and complexity of training a large vision model from scratch. We landed on combining DINOv2 with classical ML classifiers, and the results have been great.

- DINOv2 requires no fine-tuning. The pre-trained model from Meta AI converts any image into a rich embedding that already captures enough visual structure for downstream classification, right out of the box.
- The resulting feature vectors can then be used to train a classical ML classifier of your choice.

This hybrid approach gives you the intelligence of a foundation model with the agility of traditional ML. Retraining is fast and cheap as new labeled data comes in and you can iterate quickly on the classification layer without touching the expensive part. We think it's worth considering before reaching for end-to-end fine-tuning.

15/04/2026

With AI writing more and more code, security becomes more important than ever. Here's what we do at Panenco to make sure vulnerabilities never reach production:

Every PR is scanned automatically for security issues before merge. Our cloud environments are monitored continuously through dynamic analysis and pe*******on testing, not once a quarter. When vulnerabilities are detected, they are fixed automatically, not just flagged and added to a backlog. Production is observed in real time, so we know about issues before our users do. And end-to-end tests run on every release, catching regressions in CI rather than in production.
Nothing in this pipeline requires manual intervention. That's the point: security should be continuous and automated, not a checklist someone runs through before a release.

We rely on Aikido for the security scanning and remediation side, Sentry for production observability, and Playwright for end-to-end testing. Three tools that together give us full coverage from PR to production with zero manual overhead.

If you're shipping fast and want to make sure security keeps up, this setup is worth a look.

09/04/2026

We’re proud to share our latest success case with Thomas More. For this collaboration, we built an intelligent assistant that transforms company datasets into a real-time, conversational guide. This brought career opportunities directly to the student’s device instead of keeping them locked in floor plans. The impact?

- Efficiency: Students could locate companies on the floor instantly without downloading an app.
- Intelligence: Real-time answers regarding sought-after profiles, location information and company offerings.
- Accessibility: Offering students an "on-the-go" company guidebook, available whenever and wherever they need it.

Many thanks to Leo Schoeters and Wouter Lutin for the trust. We’re excited to keep innovating together to offer students better accessibility on their career navigation journeys.

03/04/2026

When testing our AI pipelines, we were previously hitting three major blockers: it can be slow, expensive (burning tokens), and non-deterministic (different outputs each run). We've recently been using PollyJS to solve all three of these problems because:

- It records API responses the first time, then replays them instantly in subsequent runs
- The full AI workflow now runs offline in milliseconds which makes it perfect for CI/CD
- Once you capture the golden path, your recordings live in git. There are zero LLM API costs and consistent deterministic outputs

PollyJS was primarily built for mocking your backend when doing frontend development, but a backend calling another backend is almost the same thing from a networking POV. A creative use case of an existing tool we'd say! If you're building with AI and watching your test suite and token spend explode, this is definitely worth a look.

Adres

Hyphen One Kolonel Begaultlaan 1A/51
Leuven
3012

Telefoon

+32497863605

Website

http://www.panenco.com/

Meldingen

Wees de eerste die het weet en laat ons u een e-mail sturen wanneer Panenco nieuws en promoties plaatst. Uw e-mailadres wordt niet voor andere doeleinden gebruikt en u kunt zich op elk gewenst moment afmelden.