Real volume.Real businesses.Real production.
Most of what we build runs inside operations that aren't public. The names stay private. The numbers, the scope and the constraints don't.
- 01Simulation Concours: a near-leader in under two months.1,050papers
- 02A freight forwarder's counter, warehouse and books, live five days after signing.5days
- 03A restaurant system across two countries, live since 2020.0churn
- 04Two chauffeur companies whose sites quote, take payment and dispatch.155pages
- 05An AI layer inside an ERP that serves 70+ companies.$5M+a week
- 06The whole backbone of a specialty perfume retailer.$200Ka month
Simulation Concours: a near-leader in under two months.
mock exams sat over five Thursdays
Scope
Design, build and run a prep platform for the Belgian medical entrance exam: courses, question bank, PDFs, a weekly mock exam, payments, and the content pipeline behind all of it.
Context
Students get one attempt a year, prep courses cost thousands of euros, and the exam was seven weeks away. We made the library and the Thursday exam free, and sold one pack for the rest. By exam day it was a near-leader: among the students preparing for the exam, the Thursday paper had become an institution.
What we built
- 01
An agent farm on one server
One process runs a bounded pool of headless agents, pauses at 90% of the usage window, resumes at the reset and never pays twice for finished work. A live dashboard shows every agent and the exact prompt it was given.
- 02
Gates before anything ships
Parsing, unique ids, a LaTeX round-trip, at least 99.5% of the source words kept and every formula identical. One source feeds the web pages and the watermarked PDFs.
- 03
Blind solvers and a refuter
Two agents answer each paper from a copy with the key removed; a third, with the key, tries to refute it. A script also checks that the right answer isn't suspiciously often the longest one.
- 04
The Thursday exam
80 original questions scored on the official scale out of 160, open from 8:00 to midnight, on screen or on a printed booklet. A ranked report the next morning, with three revision sheets picked from the paper.
- 05
Payments and access
One-time codes instead of passwords, Stripe with Bancontact and Apple Pay, and every page and PDF checked on the server before it's served.
accounts in seven weeks
students signed up for the Thursday paper
of opened checkouts paid
Launched in July, an institution among candidates by August. And every paper went out checked.
A freight forwarder's counter, warehouse and books, live five days after signing.
from signed contract to production
Scope
Replace a custom Odoo module and website with one system: the counter, the warehouse scanner, the label printer, invoicing and the export to accounting.
Context
A Brussels company sending parcels to Cameroon, Rwanda, Burundi and the DR Congo. The contract is dated 29 July 2026, the delivery note 3 August. Their history had to come across, books included.
What we built
- 01
The counter
Files, parcels, quotes and payments. Each parcel is billed on the higher of real and volumetric weight, with negotiated rates and credit notes when something changes.
- 02
Silent labels
A Windows app at the counter prints barcode and QR labels on the thermal printer without a dialog, and diagnoses the printer over the network when it doesn't answer.
- 03
The warehouse scanner
A phone app scans parcels into suitcases, pallets and containers, shows the weight against the limit, and flags a parcel bound for another destination.
- 04
Invoices and accounting
Numbering that continues the old Odoo series and can't be edited, receipts archived, notifications by email and WhatsApp, and a numbered export for the accountant.
- 05
The migration
59,453 records from 156 Odoo models exported, and 163 accounting entries moved to a newer Odoo without a cent of difference. Every clean-up in production is rehearsed and undone before it runs for real.
records out of Odoo
accounting entries, €0.00 difference
surfaces: counter, warehouse, print station, public site
Live in five days. Still shipping two months later.
A restaurant system across two countries, live since 2020.
four restaurants, still running it
Scope
An end-to-end operational system for a hospitality group: multi-device, offline-first, with owners online.
Context
Four restaurants in two countries, where the internet is a sometimes thing. That constraint shaped every decision: the system runs whether the connection does or not.
What we built
- 01
Central Linux POS
A small headless mini-PC in each restaurant holds the source of truth on the local network. It survives outages and runs the floor.
- 02
Ordering on staff phones
Servers take orders at the table and send them straight to the kitchen and the bar, across devices and floors.
- 03
Ticket routing
Each ticket prints where it belongs, bar or kitchen, based on the product. Nobody sorts tickets by hand at peak.
- 04
Real-time stock
Stock counts down with every order, on every device. The end-of-night count is already done.
- 05
Offline-first by default
Orders, payments and prints all run on the local network. The internet is a way to sync, not a requirement to work.
- 06
Cloud sync and owner dashboard
The day syncs to the cloud in batches. Owners follow revenue, recipes and performance across every venue, from anywhere.
One of the systems we're proudest of. Not the biggest. The one that's still running, five years in, in exactly the conditions it was built for.
Two chauffeur companies whose sites quote, take payment and dispatch.
in Google's index by June, up from 19
Scope
Sites, booking engines and back offices for two premium chauffeur companies: live quotes, online payment, driver dispatch, and price grids the owners edit themselves.
Context
Airport runs, corporate accounts and festival weekends, booked by travellers, assistants and hotel concierges, in French, Dutch and English at the very least.
What we built
- 01
One quoting engine
The server geocodes both addresses, checks festival fares and the fixed airport grid by radius, then falls back to the road distance. The same code answers the live quote and the checkout.
- 02
Checkout that trusts nothing
Stripe for one company, Mollie for the other. The amount is always recomputed on the server, the booking is saved before payment, and refunds go out from the back office.
- 03
Dispatch by link
The office picks a driver and a vehicle and sends the ride by email or WhatsApp. The driver answers from a one-time link and only sees their own fee.
- 04
Five languages, localized pages
Every page in French, Dutch, English, German and Spanish, generated pages per route, city and airport, structured data, and cookie-free analytics of our own.
languages on one booking flow
localized addresses in the sitemap
pages generated in three languages for the second company
Quote, payment, dispatch. The price is decided once, on the server.
An AI layer inside an ERP that serves 70+ companies.
of customer operations through the layer
Scope
Architect and ship the full AI layer of an existing ERP, end to end.
Context
We didn't build the ERP. The team behind it brought us in because their customers needed AI that runs on real production data, not a chat window bolted on top. They trusted us with the architecture, the build and the launch.
What we built
- 01
Cleaning and ingestion
Each customer's production data, prepared for retrieval in the real shape of their business. Edge cases are handled in code, not handed to the model.
- 02
Vector database and retrieval
Retrieval shaped around the ERP's data model, with vectorization tuned per data class. Queries are reconciled with the source of truth before anything is generated.
- 03
Caching
The system never pays twice for the same thought. Vectors are batched and persisted, so cold starts stay predictable under load.
- 04
Local language models
Self-hosted where data, latency or cost rule out an API. Chosen for what the system needs, not for ideology.
- 05
Mathematical validation
Anything that touches finance, operations or stock is checked against the underlying numbers before it leaves the system. The model proposes. The system checks.
- 06
Reports on request, in chat
A manager asks for a custom report in plain language. The system produces it on their own data: formatted, validated, reconciled.
The result is an automated manager that 70+ companies trust with $5M+ of operations every week. Not an API wrapper. A system.
The whole backbone of a specialty perfume retailer.
in operations on the system
Scope
Design and ship the entire operational stack from scratch: internal ERP, custom POS and a customer ordering app.
Context
Off-the-shelf POS, ERP and ordering tools didn't talk to each other, and none of them fit the way perfume is actually sold. We built one system with three interfaces instead.
What we built
- 01
Internal ERP
Inventory, suppliers, purchase orders, stock movements and staff, on one backend and one schema. Built for the people who use it every day.
- 02
Custom POS
Designed around the real choreography of a perfume sale: sampling, several items, gift wrapping, recognising the customer.
- 03
Customer ordering app
Customers order, staff fulfil, stock updates in real time. Same source of truth, no overnight reconciliation.
- 04
One backend
The three interfaces share one schema and one source of truth, so the warehouse, the till and the phone always agree.
- 05
A stock count that works offline
Since October 2026, the stock is counted on phones that keep working without a connection: every tap is queued, sent in batches and never counted twice. 2,311 products from 102 brands.
One backend. Three interfaces. One source of truth. The point was never to ship three apps. It was one system that behaves like one, from the warehouse to the till to the customer's phone.
Smaller systems.Same principle.
Beyond these six, we've shipped smaller operational systems: internal dashboards, custom workflows, integrations between tools that never talked to each other. Different scopes, same rule: software that runs the business, not software that sits next to it.
Home healthcare
Brussels
Live
Site and back office, with analytics that store no cookies and no IP addresses. A model drafts the articles; a person publishes them.
Two shops, one owner
Retail
Launching
Handheld tills that print receipts designed in the browser and update over the air, with one loyalty balance across a menswear boutique and a fishmonger.
A recruiting agent
Brussels
Demo live
The model picks a level for each criterion and quotes the profile; plain code adds up the score. Nobody is rejected automatically.
Systems we keep running.
Active retainers with several companies: web presence, internal tools and monthly evolutions. Four production websites shipped for retail and hospitality along the way.
The kind of relationship where nobody has to re-explain the business every quarter.