Heureka Labs is a technology studio from Hamburg. A network of senior experts that takes ideas from strategy to production, for clients and for ourselves.
01 client work→
Alongside your team
Before an investment, when an LLM system has to get better and cheaper, or when an idea needs its first product. Small senior teams that write the code themselves.
Statista · case study
02 product→
AI Router
Picks the best language model for every single request, by quality, cost and latency. The model calls stay on your own provider accounts.
airouter.io · in production
03 research→
Our own experiments
Knowledge graphs as memory for AI agents, an agent that steers a robot, and a Jetson computer in the stratosphere.
Engineers, data scientists and product people with many years of building and running products. Most of them have worked together before, so a new team starts working on day one.
the setup
Sized to the request
A due diligence or a strategy question needs a few experienced minds for a short time. An MVP needs a small team that builds it, right next to yours.
founder
Matthias Lau
Developer and founder. Builds AI systems, robots and AI Router, and talks about it at code.talks, OOP and the Digitale Leute Summit.
“It was such a pleasure working with you guys not only because you overdelivered on professional side, but because of your very solution oriented approach.”
Robert von Marnitz · Chief of Staff, TYTAN Technologies
Answer quality, latency and cost of an LLM application on Statista's own data, measured and improved together with their team.
VERBUND · AI agents
Agents for clean data
An agent system for large transactional databases: it learns to navigate the data and the business context, then writes monitoring queries that prevent errors.
logistics · data
CO₂ per logistics route
Data pipelines for cleaning and enrichment, CO₂ algorithms per logistics route, and a prototype ready for market testing.
investor · due diligence
AI companies, assessed
Independent assessments of product, architecture and team of AI companies, ahead of investment decisions.
client voice
“The team of Heureka has not only delivered a project in quality, budget and time. Explicitly, the team's cognitive power has paved the way to an automated, data-driven solution that maximally satisfies our customer needs.”
Robin Rohrmann · on a logistics project; today Co-Founder and Managing Director, WaveSix
software maker · agents
Coding agent maturity
How the teams work with coding agents today, and a prioritised path to the next level, presented to the supervisory board.
finance · data
Value in existing data
Years of contract and transaction data, analysed for where new products and services create value.
consumer goods · MVP
Packaging marketplace
An MVP that ingests complex packaging data, manages demand requests and matches suppliers with requesters.
Statista asked us to improve quality, performance and cost of their LLM-based research assistant. A team of five worked with Statista's own team: metric-driven experiments on vector stores, indices, reranking and retrieval.
−65 %operating cost
+140 %answer quality
−10 %average latency
client voice
“Matthias and his team were a stroke of luck for us: the combination of AI expertise in breadth and depth, the very systematic evidence-based approach and the immediate fit with our internal team on a human level is rare.”
Dr. Ingo Schellhammer · CTO, Statista
on stage
On stage with Statista's CTO
From Search Results to Insights: Learnings from Statista's GenerativeAI Journey. With Ingo Schellhammer at the Digitale Leute Summit 2024, and again at OOP 2025.
AI Router is our own product. It picks the language model for every single request, in milliseconds, so you stop paying premium prices for simple work.
why routing pays off
There is no one model that fits all
low-cost modele.g. $0.20 per million input tokenspremium modele.g. $5 per million input tokenssimple request✓right answerbest fit✓right answer25× the pricecomplex request✗failswrong result✓right answerbest fit
One model for everything means overpaying on simple requests or failing on complex ones. AI Router predicts quality, cost and latency of every model you allow and sends each request to the cell that fits.
cost
Your accounts
Every model call runs on your own provider accounts. Your negotiated prices keep applying; AI Router charges for the routing only.
control
Your rules
You choose the allowed models and how quality, cost and latency are weighted. Without clear evidence, your default model answers.
privacy
Your data
In full privacy mode, only an embedding of the prompt leaves your environment. The prompt itself never does.
pilot
Try it on your own traffic
A pilot runs 30 to 45 days on your real requests. Afterwards the dashboard shows what AI Router saves you, and which models would save even more.
“With AI Router, we've completely avoided LLM downtime while seeing our costs steadily decrease and quality improve - all without any effort on our side.”
Agents fail in real projects because they lack context. We build knowledge graphs from emails, meeting transcripts and documents that agents can reason over.
A styrofoam box with an NVIDIA Jetson, an action cam, a second camera, a datalogger, GPS and a radio module, tied to a weather balloon. Launched from a garden near Hamburg on 30 April 2022, with a permit from the aviation authority. One shot: no SSH, no reboot button.
From my garden to the stratosphere in 4 minutes. The flight replayed in our own Flight Control dashboard, with the data recorded on board. Sound on for the wind, and don't miss the burst at the end.YouTube ↗Playing loads the video from YouTube (Google). Privacy
what went wrong, and what held
01 launchwent wrong
The first ten seconds
The balloon did not go straight up. Right after release, the wind slammed the box into the balcony of the house next door. Then it climbed.
0:09 in the video above, just before the impact
“I still can't bear to look at 0:10.” Matthias Lau, founder, when he posted the video
02 temperatureheld
Freeze, or cook in its own insulation?
Batteries hate the cold, so everything sat in a styrofoam box. But insulation also traps the Jetson's heat. We tested in a freezer at −18 °C and kept the lid open until lift-off. Inside, it never dropped below freezing, and the Jetson never got warmer than 32 °C.
outside sensor, lowest −50.5 °Cinside the box, lowest +1.4 °C
03 radio linkwent wrong
The live link died at 2,027 metres
Mission control on the ground (FastAPI, MongoDB) and a live dashboard (Vue.js) waited for data over LoRa. Nine minutes in, they went dark. On board, watchdogs kept every process alive, and the Jetson recorded the rest of the flight on its own.
altitude with live data on the groundon board only
04 recoveryheld
Will we ever see it again?
After the burst the box fell 7 km in the first two minutes and landed 45 minutes later, 14 km from the garden. The GPS led us there, and every reading came back with it.
