From raw data to software that works.

I am Koen Dercksen, an AI engineer and full-stack developer. Most of my projects start with raw data, for example emails, reports, articles or web pages. I make software that changes these data into information that your team can use.

Tell me about your project See my work

Application email

Good afternoon, my name is Mark de Vries, born 14-02-1984, Kerkstraat 12, Utrecht. I have run De Vries Klussen since 2019 and would like to lease a Volkswagen ID. Buzz Cargo of €54,000 over 60 months. Last year our turnover was about €310,000.

Pseudonymise before the model sees it
My name is ‹PERSON_1›, born ‹DATE_1›, ‹ADDRESS_1›. I have run ‹COMPANY_1› since 2019 …
Extract the facts
Company age
7 years
Vehicle
Electric van
Lease amount
€54,000
Turnover
€310,000
Apply the policy rules
- Company older than 3 years
- Lease amount below 25% of turnover (17%)
- Credit registry check is not complete
Ready for the reviewer, with one open check

These examples come from real projects. The data in the examples is not real.

Long-term clients

I work with some companies for a longer period. I am part of their team, and I build their AI platforms from the first version to production.

Role: Tech lead
Since 2026

Trackbase

Trackbase shows what AI assistants tell customers about a brand.

Each day, the platform asks ChatGPT, Claude, Gemini and other assistants the questions of customers. It shows if the assistants mention a brand. It also shows the position against competitors and the sources that the assistants use.

Scheduled prompt runs on more than 10 AI engines, with analysis of mentions, positions, competitors and citations.
Analytics for AI crawlers and AI referrals, from a tracking script and CDN logs.
An insights engine that changes raw model output into recommendations with a score.
Multi-tenant accounts for agencies, subscription billing and an MCP server with OAuth.

Go to trackbase.ai

Role: CTO
Since 2024

World Brain Scholar

World Brain Scholar makes AI software for academic publishers.

The main product is Eliza, an AI assistant for peer review. Eliza helps editors and reviewers with the analysis of manuscripts and the review process.

I designed and built Eliza, and I put it into production.
A pipeline reads each uploaded paper, extracts the metadata and calculates embeddings. It also adds data from OpenAlex.
Eliza gives support to reviewers and editors and analyses each manuscript.
As CTO, I am responsible for the technical strategy and the product development.

Go to world-brain-scholar.eu

Client projects

My clients usually have a large quantity of raw data. Their teams use many hours on it each week. Here are six recent projects. I do not show the names of the clients.

Source data
AI step
Result

B2B supplier
2026

Expert blog articles in many languages

The client wanted regular expert articles on its website, in many languages. A blog engine selects a topic and researches it on the web. Then it writes the article with sources, examines it and translates it. It also makes the images and sends the article to the CMS. An editor approves each article before publication.

I improved the engine after the first version. My work is the source research, the scale and the delivery to the CMS.
Each article has references that the reader can check. The engine stops an article that has no sources.
Validation steps examine the structure, the brand rules and the factual claims before the delivery.
A delivery queue sends each article to the CMS. When the CMS gives an error, the queue tries again.

Topic profileWeb sources
Write the article with sources
Examine facts and structure
Translate and make images
Article in the CMS, ready for approval

Car-lease company
2026

First assessment of lease applications

Before this project, a person read and examined each application. Now the system reads the dossier and examines the credit and business registers. Then it makes a first assessment for the reviewer. The language models do not see data that can identify a customer.

The system replaces personal data with pseudonyms before a model sees the data. Each re-identification goes into a log.
A fixed set of rules makes the decision. These rules are separate from the language model, and an auditor can examine them.
Reviewers change the policy rules themselves. The system keeps a full history of each change.

SharePoint dossierCredit and business registers
Pseudonymise
Extract facts
Apply policy rules
First assessment for the reviewer

Investment company
2026

Lead enrichment and lead scores

Each morning, the sales team used many hours to research companies. Only then did they know which companies to call. Now the platform collects public data for each company. A language model researches the company on the web. Then the platform gives each lead a score and a short brief with sources.

The platform has connectors to business registers, insolvency records, public procurement data and news.
A background queue enriches and scores new leads automatically.
The team examines each brief. Then it sends the approved leads to the CRM.

Business registerProcurement recordsNews
Enrich
Web research with sources
Score and brief
CRM export

Financial services company
2025

Automatic news checks for all clients

Before this project, the staff searched the news for each client manually. Now a scheduled task collects articles each night. It matches the articles to the client list and classifies them. In the morning, the team sees only the important items, for example an insolvency.

Each night, the task collects articles from many news sources.
The task identifies insolvencies and other important events. It does not show routine news.
The team now reads one short list each morning. Before, the team searched for many hours.

News sources, each night
Match to the client list
Classify the event
Morning list of important items

Financial services company
2025

A private AI workspace with many agents

The team asked for AI assistants that do not send data to a public chat application. Now the team has a self-hosted chat environment with specialised agents. A separate API layer records the use and keeps answers in a cache.

The self-hosted chat environment has agents for different tasks.
An API layer between the agents and the model providers records the use and caches the answers.
The system operates on the infrastructure that the client already has.

Questions and documents from the team
Specialised agents
Layer for logs and cache
Answers in the environment of the client

Investment company
2025

The client sends a newsletter to its business contacts. A general opening does not get much attention. Now each contact receives a short, personal opening line. The line refers to a recent LinkedIn post of the person or the company, or to the last email contact.

The tool collects the contact list and the company data from the CRM.
It finds recent LinkedIn posts of each person and each company.
A language model writes one or two sentences per contact, in the tone of the client.

CRM contact listLinkedIn postsLast email contact
Find a relevant topic
Write the opening line
Personal opening in each newsletter

Own products

I also make and operate my own products.

evoicely

Valid e-invoices from your Stripe invoices.

Business customers in Germany and Belgium must receive structured e-invoices. A PDF is not sufficient. evoicely connects to Stripe and makes the correct e-invoice for each finalised invoice. It examines the e-invoice with the official validator. Then it sends the e-invoice.

Go to evoicely.com

once

Send a secret through a link that opens one time only.

Put a password or an API key in the field and send the link. Your browser encrypts the secret before it sends it. The server keeps only the encrypted data. When the recipient opens the link, the server deletes the secret. once is free, and an account is not necessary.

Go to once.avirai.com

Recent articles

Short articles about AI, data and security for teams that want to use AI in their work. Each article has sources.

28 October 20264 min read

LiteLLM supply chain attack: what to do if you installed 1.107.0

On 21 October 2026 a malicious LiteLLM release stole cloud and AI credentials. Find out if you are affected, clean the host and rotate the right keys.

Security
LLMs
Supply chain

28 October 20264 min read

Use LLMs on personal data without sending personal data

Pseudonymise documents before a language model reads them. A practical pattern for GDPR-safe extraction from emails, dossiers and applications.

Privacy
GDPR
LLMs

All articles

Hello, I am Koen.

For my PhD research at Radboud University, I worked on natural language processing. My subject was radiology reports. I made these reports easier to understand for patients.

Now I am a senior engineer at Plain Medical, where I work on medical text. I am also the CTO of World Brain Scholar.

With avirai, I do a small number of client projects at the same time. I do each project from start to end. First I find where AI helps. Then I make the system and operate it in production. You talk directly to the person who writes the code.

Location: Nijmegen, the Netherlands
Languages: English and Dutch
Education: PhD research in NLP, MSc Data Science, BSc AI

Tell me about your project

Write a short description of the problem. Tell me which type of organisation you are and which tasks use too much time. Also tell me what you tried before. I will send a reply by email.

You can also send an email to info@avirai.com.