AI Engineer and full-stack architect.
I take AI beyond the prototype.
For more than 15 years I have been building and running software: applications, data platforms and cloud infrastructure, from the architecture draft through to live operations. Today I mostly work on AI systems — LLM applications, RAG, agents. The hard part is rarely the model; it is data quality, integration with grown systems, evaluation, cost and operations. That is craft from my fifteen years before that.
Porsche
Lidl
KUKA
Agilent Technologies
GC Gruppe
Maico
Roca
Condor
EnBWFive projects from recent years.
Projects where I did the development, owned the architecture or led them technically — from AI agents and data platforms to complex full-stack applications. I don't attribute these projects to any client here, because most of them are under NDA.
Website builder, thousands of sites in production
A platform that generates customer websites automatically from the sales process and simplifies their ongoing maintenance.
Platform architecture, technical leadership and development of the core editor and automation features.
Thousands of websites with 10 to 30 subpages each in production. New sites and many customer changes are created without a development step in between.
Technical details
Drag-and-drop editor with nested components, responsive breakpoints, live preview, undo/redo with versioning, templating, theming and media management on a headless data model built with React, Next.js and MongoDB. A supplementary AI feature applies changes phrased in natural language within the controlled design system. On top of that, a sales configurator, provisioning from Salesforce via n8n, plus deployment and operations on Kubernetes.
Agentic BI: querying company data in natural language
A complete BI system that consolidates distributed operational data into a shared data foundation. Alongside fixed KPIs and dashboards, an AI agent also answers individual questions in natural language.
Concept and full development of the ETL pipelines, data preparation, KPI logic, Metabase analytics and the BI agent.
The system serves as the shared basis for complex operational and strategic decisions today. Recurring questions are answered through Metabase; new ones can be investigated with the agent without having to build a dedicated report first.
Technical details
The operational data sits in several local MongoDB and PostgreSQL systems. Various ETL pipelines pull it in via Airbyte, prepare it for analysis and consolidate it into a central PostgreSQL data foundation. Metabase then covers the recurring KPIs, dashboards and interactive analytics on top of it.
For questions that are not yet prepared as a report, an AI agent complements the classic BI interface. It translates naturally phrased questions into SQL in the background, queries the shared data foundation and presents the result in an understandable form. That way the system covers both stable standard reporting and ad-hoc analysis.
Agent-ready CRM: an AI agent that knows when to ask
A CRM in which an AI agent doesn't just suggest text but works on the same business processes and data as the team.
Concept and development of CRM, ERP synchronisation, agent logic and the controlled knowledge base.
Routine cases run automatically; uncertain decisions land with the team, with reasoning attached. The ERP remains the system of record.
Technical details
Multi-tenant CRM built with Next.js, TypeScript and PostgreSQL, connected to weclapp through a bidirectional sync engine. Contacts, quotes, orders and email threads stay in sync. The agent reviews cases, detects duplicates, carries processes forward and answers inquiries from a second brain. Below a confidence threshold it asks; an auditor agent controls its work and the intake of new knowledge.
Data platform with GenAI: bulk operations across Salesforce, ERP and databases
Complex bulk changes in Salesforce, an ERP and databases in natural language — with a preview instead of a blindly executed script.
Concept and development of execution engine, sandbox, connectors and approval process.
No generated change reaches a real system before a human has confirmed the before/after diff. Rollback and tests add a further safety layer.
Technical details
A language model plans the operation and generates its logic. It first runs read-only in an OS-isolated sandbox with no network access; writing happens only after the diff is confirmed, with rollback to the previous state. Target systems are connected through connectors whose allowed write fields are approved by a human. FastAPI backend, React frontend, tenant isolation and encrypted credentials.
Client portal for a law firm: read case files, annotate them, work on them together
A portal where clients and the firm view, annotate and work on the same documents — even when the case file arrives as a scanned PDF.
Concept and development of the portal backend, the document pipeline and the viewing and annotation features.
Clients and the firm work on the same document instead of on email attachments. Scanned case files are searchable, not just viewable.
Technical details
Java and Spring Boot backend with a processing chain that runs incoming documents through OCR into searchable text and prepares them for display. On top of it a PDF workspace in the browser: page-accurate viewing, highlights and annotations, comments on the document, plus shared access for clients and the firm through graduated permissions.
More projects and details on request.
Four things I'm good at.
The short version. I'd rather go into detail on a concrete project.
AI architecture & GenAI in production
I design AI systems so they survive the prototype: chatbots with their own knowledge base, RAG setups with vector databases (Qdrant), AI agents, prompt engineering and ML pipelines. Plus the questions AI projects actually fail on: data quality, connecting to grown systems, evaluation, cost, data protection (GDPR) and operations. Most of it runs in clients' daily business today, not on slides.
Architecture & system design
Translating business requirements into solid technology: system boundaries, technology decisions, data flows. I learned that craft back when architecture had to work entirely without AI — and it is exactly what decides today whether an AI system holds up. I'll also tell you when the simple solution is enough.
Full-stack development
Angular or React with TypeScript at the front, Node, Java or Python behind it, REST and GraphQL in between, PostgreSQL or MongoDB underneath. Plus data visualisation when numbers need to become understandable, and automated tests so that changes are not a gamble. And no fear of someone else's code.
Cloud-native, DevOps & operations
I've been running my own Kubernetes clusters for years, on Hetzner among others, and work with Azure and AWS on client projects. CI/CD, monitoring and backups are part of the project.
I join a team as a developer or as an architect — for classic applications just as much as for AI initiatives. The four areas can be booked individually.
Who I am.
I am Thoren, a graduate engineer in software development, and I have been building software by hand for more than 15 years: front ends, back ends, data platforms, plus my own production cloud infrastructure for years now. I have ridden every wave, from Java and C# through web and cloud to AI today.
Today my focus is on AI. The hard part of an AI project is rarely the model. It sits in the data, in the interfaces to grown systems, in evaluation, cost, data protection and operations. That is exactly where I am at home, because I spent years building systems like these myself. My own working day is AI-supported too.
I join a team to strengthen it, not to be a one-man show. Most of the time I am the technical link between the business side and development: I sit down with the business people until it is clear what is really needed, and I explain technical decisions so that non-technical people can help decide. Depending on the project I set the architecture, support less experienced developers, or simply sit in the code myself until it runs.
And when I think something is a bad idea, I say so. That saves time and money on both sides.
What I write about.
On Medium I publish guides from the engine room: n8n, Kubernetes, Superset, scraping. I set everything up myself before I write about it.
How to install this Free Open Source Zapier-Clone Solution with 42k stars on GitHub — n8n
Read on Medium → June 2024Why Streamlining B2B Workflows with Apache NiFi is Fantastic — And Why We Still Switched
Read on Medium → June 2024Robot Automation: Run any Git Repository as a Robot in Kubernetes with an easy Configuration UI —…
Read on Medium → June 2024Stabilise your Web-Scraper for a 100% Success-Rate with this Cool Tool — n8n.io
Read on Medium → May 2024Query Data from MongoDB in Apache Superset with the Help of Apache Drill. (Full Tutorial)
Read on Medium → May 2024This Open Source Dashboard Solution With 59k Stars on GitHub astonished me — Apache Superset
Read on Medium → May 2024Upgrade your super low price kubernetes cluster in hetzner in some easy steps
Read on Medium → May 2024Added a free VPN to my Selenium Scraper Docker Container.
Read on Medium →All articles on Medium →
Wind, water, family.
Family man, happiest wherever there's wind and water. Kitesurfing has been my balance for years.
I organize my projects remotely and reliably, independent of location.
Got a project in mind? Write me.
Email is fastest. I usually reply within a day, even if the answer happens to be that I'm not the right fit.
thoren@hellothoren.com