Menu
Nederlands

From raw data to software that works.

I'm Koen Dercksen, an AI engineer and full-stack developer. Most of my projects start with messy data: emails, reports, articles, web pages. I build the software that turns it into information your team can work with.

Tell me about your projectSee 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.

  1. Pseudonymise before the model sees it

    My name is ‹PERSON_1›, born ‹DATE_1›, ‹ADDRESS_1›. I have run ‹COMPANY_1› since 2019 …

  2. Extract the facts

    Company age
    7 years
    Vehicle
    Electric van
    Lease amount
    €54,000
    Turnover
    €310,000

  3. Apply the policy rules

    • Company older than 3 years
    • Lease amount below 25% of turnover (17%)
    • Credit registry check still pending

    Ready for the reviewer, with one open check

Based on real projects. The data in these examples is made up.

What I can help with

I build all kinds of web applications, with a focus on applying AI to real business problems. Most projects start with one of these.

Long-term clients

With some companies I work long-term. I'm part of their team, and I build their AI platforms from the first version all the way to production.

  • Role: Tech lead
  • Since 2026

Trackbase

Trackbase shows what AI assistants tell customers about a brand.

Every day, the platform puts customer questions to ChatGPT, Claude, Gemini and other assistants. It shows whether they mention a brand, how the brand ranks against competitors, and which sources the assistants rely on.

  • Scheduled prompt runs across more than 10 AI engines, with analysis of mentions, rankings, competitors and citations.
  • Analytics on AI crawlers and AI referrals, from a tracking script and CDN logs.
  • An insights engine that turns raw model output into scored recommendations.
  • 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 builds AI software for academic publishers.

Its main product is Eliza, an AI assistant for peer review. Eliza helps editors and reviewers analyse manuscripts and get through the review process.

  • I designed and built Eliza and took it to production.
  • A pipeline reads every uploaded paper, extracts the metadata, computes embeddings and adds data from OpenAlex.
  • Eliza supports reviewers and editors with an analysis of each manuscript.
  • As CTO, I'm responsible for the technical strategy and product development.
Go to world-brain-scholar.eu

Client projects

My clients usually have a lot of raw data that costs their teams hours every week. Here are eight recent projects, with the client names left out.

  • Source data
  • AI step
  • Result
  • B2B supplier
  • 2026

Expert blog articles in many languages

The client wanted a steady flow of expert articles on its website, in many languages. A blog engine picks a topic, researches it on the web, writes the article with sources, checks it and translates it. It also creates the images and delivers the article to the CMS. An editor approves every article before it goes live.

  • After the first version, I worked on the engine's source research, its scale and the delivery to the CMS.
  • Every article has references readers can check. An article without sources doesn't get through.
  • Validation steps check the structure, brand rules and factual claims before delivery.
  • A delivery queue sends each article to the CMS and retries when the CMS returns an error.
  1. Topic profileWeb sources
  2. Write the article with sources
  3. Check facts and structure
  4. Translate and create images
  5. Article in the CMS, ready for approval
  • Car-lease company
  • 2026

First assessment of lease applications

Someone used to read and check every application by hand. Now the system reads the dossier, checks the credit and business registers, and prepares a first assessment for the reviewer. The language models never see data that could identify a customer.

  • Personal data is replaced with pseudonyms before any model sees it, and every re-identification is logged.
  • The decision comes from a fixed set of rules, kept separate from the language model, so an auditor can inspect them.
  • Reviewers adjust the policy rules themselves, with a full history of every change.
  • The reviewers save an estimated 60 to 80% of the work on each application.
  1. SharePoint dossierCredit and business registers
  2. Pseudonymise
  3. Extract facts
  4. Apply policy rules
  5. First assessment for the reviewer
  • Investment company
  • 2026

Lead enrichment and lead scores

The sales team spent hours every morning researching companies before they knew who to call. Now the platform collects public data on each company, a language model researches it on the web, and every lead gets a score and a short brief with sources.

  • Connectors to business registers, insolvency records, public procurement data and news.
  • A background queue enriches and scores new leads automatically.
  • The team reviews each brief and sends the approved leads to the CRM.
  1. Business registerProcurement recordsNews
  2. Enrich
  3. Web research with sources
  4. Score and brief
  5. CRM export
  • Software company
  • 2026

A job queue for shared GPU machines

A team of engineers and researchers shared a handful of GPU machines and used a group chat to agree who got which one. Now everyone submits a Docker job to a queue, which starts it on a machine with free resources. A web interface shows the queue, the logs and how busy each GPU is.

  • Jobs are placed by the GPU memory they need, so small jobs can share a GPU and less GPU time goes to waste.
  • Each job runs as the person who submitted it and can only access the folders the rules allow.
  • Users can open a shell in their job. A job can also start a model server for other jobs, which shuts down once those jobs finish.
  • Coding agents can submit jobs and read the results too, through an MCP server.
  1. Docker job from a team memberFree resources on each machine
  2. Select a machine
  3. Run the job as its owner
  4. Logs and GPU use in the web interface
  • Software company
  • 2026

Local language models, only when the team needs them

The team wanted to use language models without sending data to an outside provider. But a model on a local GPU holds on to a large chunk of GPU memory, even when nobody is using it. Now the team has one API for all local models. An agent on each GPU machine starts a model on the first request and stops it after an hour without requests, which frees the GPU for other work.

  • One OpenAI-compatible API for all local models, which team members connect their chat and coding tools to.
  • A dashboard shows which models are running, and admins can start or stop any of them.
  • The models run on GPU machines the company already owns. No data goes to an outside provider.
  1. Request from a chat or coding tool
  2. Shared API
  3. Start the model if needed
  4. Answer from the local model
  • Financial services company
  • 2025

Automatic news checks for every client

Staff used to search the news for each client by hand. Now a scheduled job collects articles every night, matches them to the client list and classifies them. In the morning, the team sees only what matters, such as an insolvency.

  • Every night, the job pulls in articles from many news sources.
  • It flags insolvencies and other important events and filters out routine news.
  • The team now reads one short list each morning instead of searching for hours.
  1. News sources, each night
  2. Match to the client list
  3. Classify the event
  4. Morning list of important items
  • Financial services company
  • 2025

A private AI workspace with many agents

The team wanted AI assistants that don't send data to a public chat app. Now they have a self-hosted chat environment with specialised agents, plus a separate API layer that logs usage and caches answers.

  • A self-hosted chat environment with agents for different tasks.
  • An API layer between the agents and the model providers that logs usage and caches answers.
  • Everything runs on infrastructure the client already had.
  1. Questions and documents from the team
  2. Specialised agents
  3. Layer for logs and cache
  4. Answers inside the client's own environment
  • Investment company
  • 2025

A personal opening line in each newsletter

The client sends a newsletter to its business contacts, and a generic opening doesn't get much attention. Now every contact gets a short, personal opening line that refers to a recent LinkedIn post by the person or their company, or to the last email exchange.

  • The tool pulls the contact list and company data from the CRM.
  • It finds recent LinkedIn posts by each person and each company.
  • A language model writes one or two sentences per contact, in the client's tone of voice.
  1. CRM contact listLinkedIn postsLast email exchange
  2. Find a relevant topic
  3. Write the opening line
  4. Personal opening in each newsletter

Own products

I also build and run a few products of my own.

evoicely

Valid e-invoices from your Stripe invoices.

Business customers in Germany and Belgium have to receive structured e-invoices; a PDF isn't enough. evoicely connects to Stripe, creates the correct e-invoice for every finalised invoice, checks it against the official validator and sends it.

Go to evoicely.com

once

Share a secret through a link that works only once.

Paste a password or API key and send the link. Your browser encrypts the secret before it leaves your device, so the server only stores encrypted data. As soon as the recipient opens the link, the secret is deleted. once is free, and you don't need an account.

Go to once.avirai.com

Recent articles

Short articles on AI, data and security, for teams that want to use AI in their daily work. Every article lists its sources.

All articles

Photo of Koen Dercksen

Hi, I'm Koen.

For my PhD at Radboud University, I worked on natural language processing for radiology reports, with the aim of making them easier for patients to understand.

Today I'm a senior engineer at Plain Medical, working on medical text. I'm also the CTO of World Brain Scholar.

I keep avirai small on purpose: a few client projects at a time, each one from start to finish. I work out where AI can save your team time, build the system and keep it running in production. And you deal directly with the person writing 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

Describe the problem in a few sentences: what kind of organisation you are, which tasks take too much time and what you've already tried. I'll get back to you by email.

You'll hear back within one working day. The first conversation is free and without obligation, and I'm happy to sign an NDA first.

Or email me directly at info@avirai.com.

I only use this information to reply to you. Read the privacy policy.