Production-Ready AI · Remote Worldwide
Production-ready AI that impresses your users
Whether it’s a prototype, a startup MVP, or AI features for software with thousands of users: I build cutting-edge, reliable and user-friendly AI solutions for B2B SaaS.
12+ years production ML · Ex-CTO PAXLY · Production AI in regulated software · CKA 2025
Where I help
Wherever your product is right now
Teams usually bring me in to build a new AI product, extend an existing one, or take a prototype live:
An AI feature is on the roadmap, but the team lacks the ML and LLM experience to build it solidly.
A prototype is winning over customers and investors, and now it needs to become a real feature or a full product.
The feature works, but cost or latency need to come down without hurting quality.
The first “AI consultant” delivered notebooks and slides, but no software you can put in front of customers.
Approach
How I work
Three clearly separated phases. After each one, you decide whether the next makes sense.
01 · 1–2 weeks
Scoping & plan
I get an overview of the project goal, the data you have, and the current state of the software. We define architecture and workflow and set priorities. At the end you have a written plan breaking down scope, cost and milestones.
02 · 2–8 weeks
Build & improve
We implement the project plan, and you see progress regularly through demos and can give feedback.
03 · Project close
Handover
Your team can run and extend the feature without me. Full documentation of the architecture and the decisions made.
Outcome
What you end up with
An AI feature that earns your users’ trust and holds up in production.
Measurable quality instead of gut feeling: evaluation and monitoring, so your team can see how users actually engage with the AI.
Automated tests that catch errors before they reach your customers.
Clean, documented code and architecture your team can keep building on.
Cases
Relevant cases
Selected projects with measurable impact — startups and client work. I’m happy to name the clients in a call.
Regulated Medical Software
For large customers, core features were so slow they were effectively unusable — users assumed they were broken. Optimized algorithms, caching and optimized database queries brought them back to life: up to 90% faster load times. Plus an AI assistant with intent classification: 95% fewer wrong answers — built to the V-model.
95% fewer wrong answers · up to 90% faster load times
TypeScript · Node.js · PostgreSQL · LLM
AdTech · Real-Time Advertising
Regression-tree model for floor-price optimization in real-time ad auctions: around +8% revenue, in the millions. Plus a data pipeline with an 80% smaller data warehouse and 75% less reporting load.
≈ +8% revenue · −80% data volume
Java · C++ · Spark · ML
Audit Data · Food Safety
Explainable anomaly detection for audit data: every flag comes with a full reasoning chain — the system tells you not just that something is unusual, but why. Decisions stay traceable and defensible for auditors.
Full reasoning chain · no black box
Python · PGM · Neural Nets
Climate-Tech · Data Pipeline
A pipeline ran documents through an expensive LLM — at millions of documents, cost explodes. Instead of naively filtering with the LLM, the LLM trains a cheap TF-IDF model that filters upfront: only ~10,000 of 3M documents reach the expensive stage. LLM costs cut by ~99%.
≈ −99% LLM cost
Python · AWS · Ray · LLM
My startups
PAXLY
ML-based packaging optimization saving e-commerce companies like Flaconi 10–15% on daily shipping costs. Built as ex-CTO, €1M+ funding, still in use today. Engineered so robustly that critical hotfixes dropped from one or two a month to about one a year.
Java · Python · AWS
Firmbase
Company search engine for the UK market with 5.5M+ records, built full-stack. AI-powered data enrichment end-to-end — ~80% relevant matches per search vs ~20% with common tools, roughly 5× more precise lead lists.
TypeScript · PostgreSQL · Python · LLM agents
FAQ
Frequently Asked Questions
How are you different from a prompt engineer?
I build complete AI features and take them live with everything that entails: architecture, database integration, monitoring, cost control and tests. Prompt tuning is a tool, not a job description. My background is 12 years of machine learning, long before LLMs were a thing: regression trees in ad tech, probabilistic graphical models for audit data, ML-based packaging optimization as a CTO.
Do you also do classic software development, not just AI?
Yes. I’m a senior software engineer with 12 years of experience across backend, data modeling, full-stack, infrastructure and AI. AI is the specialization on top, not a replacement for solid engineering. You get both from one person: a cleanly built feature with AI that actually works.
Do you build new AI features or improve existing ones?
Both. Many engagements start from zero: a feature is on the roadmap and the team lacks ML/LLM experience. Just as often, an existing feature needs to become more reliable, faster or cheaper, or a prototype needs to make the jump to a live system. In every case, the first step is a short planning session that ends with a concrete plan.
Do you also work in regulated industries?
Yes. I worked on an AI assistant for regulated medical-device software, built in a validated software-development process (V-model). If your use case is in a regulated industry, reach out directly.
Do you work remotely?
Yes. I work fully remote with teams worldwide, and I’m available on-site around Vienna. I’m in CET (Vienna), which overlaps well with European hours and US mornings.
Ready for AI that wins your users over?
Tell me about your use case in a free intro call — whether it’s new, half-finished or already live. You’ll get an honest first assessment of the biggest levers and a realistic proposal for the next few weeks.
Currently available for new projects
Book a free intro call