Services
Custom AICustom SoftwareData MigrationAI chat environment
Menu
ApproachCasesAbout usContactBook a call
Cases

No slide decks. Working software.

Everything here runs at a client or has been proven on real data. Every case has a status, so you can see what is production and what is still proof of concept. Names anonymised for client confidentiality; in a conversation we're happy to share more, including what didn't work.

  • 6 in production
  • 4 as proof of concept
  • 7 sectors
Case library

Thirteen projects, from proof of concept to production

Production means: runs daily at the client. Proof of concept: working on real data, with a decision point. In development: commissioned, module being built. Proposal: scope and quote are with the client.

Production
Food & Agri · Document AI

6,800 supplier invoices from PDF into one procurement database

A tomato grower with sites in Belgium and the Netherlands had six years of invoices as PDFs, in dozens of layouts, via email and WeTransfer. A pipeline reads them automatically from SharePoint, recognises supplier and entity, extracts every line and checks the total against the printed amount. Edge cases reach the buyer as proposals, approved with one click. On top of the database: a procurement dashboard with price monitoring, categories and data quality.

6.873PDFs in the archive processed
100%checked against printed total
Production
Retail & Interior · Image AI

4,500 room impressions for a wallpaper brand, with quality control

An interior brand delivers mood images per design in eight room types and eight styles. A batch line processes up to fifty designs per run: the design is never regenerated, only cropped to wall ratio, and every impression passes a vision check for fidelity, distortion and mirrors. Plus a wall preview that places the wallpaper on a customer photo while preserving windows and doors, and a print pipeline up to 10,000 pixels.

4.564impressions produced
$0,08per impression, including QC
Proof of concept
Aviation · Document validation

Thousands of pages of maintenance documentation per aircraft engine checked

With every purchase of an engine or part, the full history must be right: every day, flight cycle and flight hour accounted for, in unstructured packages of thousands of pages. Errors hide in details humans miss: a wrong serial number, a non-matching period, a missing signature. We designed a pipeline that classifies documents, extracts key fields, builds a timeline per component and bundles deviations for the engineer. Their own chat environment on the client documentation is already running.

5 → 1days of review per package (target)
Humanreviews every deviation
Production
Food & Agri · Data & dashboards

Forecast versus actuals per greenhouse, daily from SharePoint

Cultivation forecasts lived in a dozen Excel files, each maintained by a different colleague. Now a pipeline reads the source files daily from SharePoint and one dashboard shows forecast accuracy per location, crop and horizon, including the late adjustments nobody used to see. An assistant in the chat validates new forecasts and flags outliers.

17locations in one overview
12 → 1spreadsheets to one source
Proof of concept
Engineering & Installation · Quote agent

From request with drawing to draft quote in brand style

A supplier of access and parking systems received requests as PDF, Excel, Word and photos. An intake reads the request with vision, matches positions against a confidential price model that never reaches the agent, and renders a draft quote as Word and PDF. Validated against real quotes to the euro. New: a dimensionally accurate 3D product visualisation generated from the calculation model.

€66.362test quote matched to the euro
Minutesinstead of days
Production
Garden & Construction · Document generation

Assembly manuals in four languages, automatically from supplier documents

A garden-timber wholesaler uploads the supplier manual and its own product sheet; the tool generates an assembly manual as a PDF in the brand's style, in Dutch, German, English and French, with a public page and QR code for the shipping label. Scanned documents go through a vision route, vector files through exact label translation.

4languages from one source
QRon the shipping label
Production
Retail & Interior · AI chat environment

Own AI environment in brand style, EU-hosted, with the best image models

A private chat environment with their own name and colours, where staff use a wallpaper designer, intake with copyright check and order flow. Models from Anthropic, OpenAI, Google and ByteDance behind one gateway, the best per task. Roles, logging, budgets and error alerts to Slack.

76days continuously live
100%EU-hosted, no training on data
Proof of concept
E-commerce · Translation platform

Translating an entire web shop with three models checking each other

The Dutch site of a timber trader is the most extensive; the German and French versions lag behind. A platform reads the sitemap, has several models translate each page independently, presents differences to a reviewer and guards terminology and links with deterministic checks. Cost and time per translation are visible; loose PDF and Word files can be added too.

3models, consensus per page
Whole siteread via sitemap
Production
Marketing & Lead generation · Data agents

Twenty-five analysis agents on millions of transactions, no SQL required

A cashback and savings platform with hundreds of thousands of members had its data in Elasticsearch, accessible only to developers. Via an MCP connection, agents now answer marketing and finance questions: churn reasons, approval ratio per network, look-alike mailing lists, provisions. A mailing agent prepares draft mailings in their own admin system.

25+reusable agents
0lines of SQL for the user
Proof of concept
Food & Agri · Image to data

Supermarket shelf photos become market data

Field workers photograph shelves in supermarkets across several countries. Vision AI reads brand, EAN, price, promotion and packaging from each photo and normalises them into one trade-marketing dataset. Five different source types, one data model.

5source types into one model
Photobecomes a structured row
In development
Construction & Interior · Quote intake

A hundred quotes a month, soon three quarters AI-driven

A builder of glass office spaces produces a hundred quotes a month from drawings, Excel schedules and emails. After the proof of concept on real projects we're building the intake into a module that prepares the calculation and connects to the existing quoting system, with a roadmap towards 75 percent automation.

100quotes per month
75%AI-driven as target
In development
Retail & Interior · Platform

An in-house visualizer replacing a US SaaS subscription

The client pays a foreign vendor over two thousand dollars a month for a wall visualizer that uses customer photos for AI training. We're building the foundation for an in-house platform on Cloud Run: deterministic pattern placement at centimetre scale, byte-exact mask guarantee, SKU catalogue and a quality-check API for customer-service photo orders.

$2.200per month of SaaS to replace
0customer photos outside the EU
Proposal
Food & Agri · HR intranet

An intranet with an AI search bar on the HR documents in SharePoint

One landing page with a large search bar that answers questions from HR documents with source references, a news carousel colleagues fill themselves and eight tiles to SharePoint and external tools. SharePoint remains the cabinet, RocketCode builds the front.

8tiles, one search bar
Sourcewith every answer

Names anonymised for client confidentiality. References available on request.

What these cases have in common

Three things that keep coming back

One process, not an 'AI strategy'

Every case started with one recurring, measurable process. Not with a tool and not with a vision document.

Data first, then AI

In almost every project, data had to be consolidated or cleaned first. Only then did AI become reliable.

A human at the controls

Every solution has a place where a human reviews edge cases. That's not a weakness, that's why it's allowed in production.

Next step

What will your first case be?

Tell us which process costs the most time. In a free scan we show what a first working version looks like.

Book a free scan
Book a call