Florian
Zeba.
I design the platforms enterprises run their data on, and I still write the pipelines that fill them. An architecture is only as good as the thing you can point at when a number is wrong.
Principal Enterprise Architect · Data & AI, working across consultancies, systems integrators and my own practice.
Career
Consulting and building, at the same time — not in sequence.
A CV flattens a career into a list, which hides the part that matters. The consultancy roles, the engineering work and my own practice were not phases. They ran together. Titles and shape below; the organisations stay off the page.
- Principal Enterprise Architect
- Data & AI Architect
- Independent practice
- Software Engineering
- Tax Technology Engineer
- Senior Associate · BI
-
Principal Enterprise Architect · Data & AI
-
Data & AI Architect
-
AI, Data & Software Engineer
-
Software Engineering
-
Tax Technology Engineer & Consultant
-
Senior Associate · Management & BI Consulting
Research
Governance you can prove, not guardrails you hope hold.
Cortex: A Fixed-Point Theory of Governed Coding Agents — co-authored with Marius-Constantin Dinu at Alpha Omega Labs. It treats an agent's validate–repair loop as a monotone operator on a lattice of satisfied requirements, so “done” becomes a least fixed point you can prove rather than a heuristic stopping rule.
Adversarial attack success
42.4% → 24.1%
Capability on single-file tasks
Parity
Pages
21
Writing
Notes on making data trustworthy.
All posts →- The Graph Is Not the Point: Building a Personal Knowledge System That Actually Teaches You Something Why building a concept map beats studying one, why the untyped wikilink throws away the information you needed, and how to layer embeddings, LLM triple extraction, and structural gap detection on top without letting the machine author your graph.
- Building a Write-Back Interface on Microsoft Fabric Apps with Rayfin A full technical walkthrough of building a governed write-back interface on Microsoft Fabric Apps with the Rayfin SDK — static hosting, Entra ID SSO, and a server-side functions proxy to an existing REST API, with no separate web app to maintain.
- Stop the Garbage Before It Lands: A Deep Look at Databricks Labs DQX A practical guide to DQX, the PySpark-native data quality framework from Databricks Labs that quarantines bad data before it reaches your gold tables.
- Multi-Agent Pipelines for Academic Figures: The PaperBanana Open-Source Implementation A technical deep dive into PaperBanana, an open-source agentic pipeline that generates publication-quality methodology diagrams and statistical plots from text.
- Stop Tinkering With Strings: A Practical Tour of DSPy and Reflective Prompt Evolution Why Stanford's DSPy framework treats prompts as compiled artifacts, and how GEPA — the ICLR 2026 Oral — outperforms reinforcement learning with 35x fewer rollouts.
- Beyond Tidy Data: How Microsoft Data Formulator Lets AI Reshape Your Way to a Chart Exploring Microsoft Research's Data Formulator, the concept-binding paradigm, and how its AI agents remove the tidy-data tax from visualization authoring.
Get in touch