01 / Position
What works.
What doesn't.
What we ship.
supadupa is an AI engineering studio in Vienna. We research how AI systems actually behave, publish what we find, and build the products that come out of it. An open standard, an open benchmark, open results. Including the ones that went against us.
02 / Stack
From the specification
to the invoice.
Six public artefacts, five layers, one house. Every line below is a link to something you can read, run or test yourself. That is the whole argument.
Standard
Open Brand Definition
openbranddefinition.org
An open specification for how brand truth is resolved, verified and recorded inside AI systems. Written by us, licensed for anyone.
v3.0.2 · 1,079 tests · CC BY 4.0 + Apache 2.0
Research
Memory utility
supabrain.space
Counterfactual ablation for retrieval memory. We removed each memory and measured what broke. The method held. Our own trained model did not beat BM25, and we said so.
23 pages · open data · code on GitHub
Benchmark
trashfire
trashfire.io
An open benchmark for AI code review. Real projects, planted bugs, encrypted ground truth, six weighted dimensions. Bring your model, get a score.
42 projects · 30+ languages · 4,200+ bugs
Infrastructure
supabrain
supabrain.dev
A context compiler for agents. Bounded evidence packages under a hard token budget, local SQLite, provenance down to the line and the hash.
12.4 s → 0.083 s · 650 MB → 140 MB
Infrastructure
supaskills
supaskills.ai
A skill API for coding agents, over REST and MCP. Every skill quality scored on six dimensions and sourced before it ships.
1,307 skills · 6 domains · live
Product
supakraft
supakraft.ai
The operational layer for brand and marketing. Brand check, content generation, RAG workbench. Where the standard and the infrastructure meet a paying customer.
EU-hosted · GDPR · EU AI Act · in production
03 / Method
Three claims.
Three receipts.
01
We measure before we claim.
Our memory research came back negative. A plain BM25 retriever beat the specialist we trained. We published the paper, the data and the method anyway.
Read the paper →02
We build the foundation, not just the app.
Open Brand Definition is a full specification with a conformance suite, a schema and an open licence. 1,079 tests, zero failures, free to use commercially.
Read the spec →03
We ship into production, not into demos.
supakraft runs on isolated EU instances under GDPR and the EU AI Act. One industrial pilot, measured over twelve months, not over one afternoon.
See the product →04 / Work
What we do
for clients.
Three engagements. Fixed scope, named deliverables, a timeframe you can hold us to.
01
AI systems and SaaS
For teams who need a system in production, not a prototype in a deck.
You get
- Architecture and running code
- Tests, CI, deployment
- Handover docs your team can act on
Typical timeframe
6 to 12 weeks to first production release
02
Agent and workflow infrastructure
For teams whose people still do work a system should be doing.
You get
- API and MCP integrations, agent layer
- An evaluation harness, so you can tell if it works
- Monitoring and cost control
Typical timeframe
4 to 8 weeks
03
Brand and knowledge layer
For organisations whose rules have to survive contact with a language model.
You get
- Knowledge audit and structured brand definition on OBDS
- Compiled runtime, brand check, governance
- A standard you own, not a black box you rent
Typical timeframe
Audit in 2 weeks, layer in 8 to 12
05 / Who
Vienna.
Small on purpose.
supadupa is the product and engineering side of a two-company setup. Kill The Dragon GmbH does brand and communication work and publishes Open Brand Definition. supadupa builds the systems that run on it. Same people, same address, two jobs.
No account layer, no delivery pyramid. You talk to the people writing the code.
- Entity
- Kill The Dragon GmbH
- Address
- Dürergasse 3 / TOP 4 / HOF, 1060 Vienna, Austria
- Managing Director
- Max Peter Jürschik
- Publications
- Research and standards →
06 / Contact
Send us the
hard one.
Tell us what has to be true in production and by when. You get a straight answer on whether we are the right shop, usually within a day.