Hi, I'm Marci,
Verifiable AI for regulated, high-stakes environments: a citation on every answer, an audit trail on every action, and a human in the loop where the regulator expects one. I design it, build it and run it, and I sit in the room when your compliance team asks how it works.
Four engagements. The right-hand column says what a third party can verify, and where.
Structure-aware chunking, hybrid retrieval with reranking, citation provenance on every answer and an immutable audit trail, orchestrated as Temporal workflows with human approval steps. A Terraform-managed Google Cloud platform of about 180 resources, reviewed plan by plan.
Every answer traces to a passage in the source document.
Human approval is a step inside the workflow, not a policy.
Case study in preparation, published with the client's sign-off.
I led the research team that took Deloitte's reusable KYC credentials to production on the KILT network: system architecture for the self-sovereign identity protocol through its mainnet launch on Polkadot, a compliant token sale, and the KYC and KYB credential that Deloitte issues on it.
Live on mainnet since 2021.
Deloitte issues the credentials.
Sources: Finextra, Deloitte on LinkedIn, KILT Protocol, and the Polkadot Decoded 2023 talk.
A four-agent coaching backend with dynamic prompt assembly, journey-gated tools, a safety mode for overwhelmed users and full tracing through Langfuse. The coach's rules moved out of code into a file the team owns.
Two weeks to days per new feature.
Zero deployments to change what the AI says.
Full write-up being moved to this site.
Airtable, n8n and the WhatsApp Business API, with an LLM memo writer wrapped in a code-level guard: every figure is rendered from a field, never generated. Readiness-gated announcements and commitment tracking, live in production.
No number in a memo comes from the model.
Live with real investors since August 2026.
Reference available on request.
Smaller engagements: email triage for a German manufacturer, 95%+ categorisation accuracy and 60% less manual handling; a support agent for an e-commerce startup, 70% less manual support work; a content pipeline for a marketing agency, 450+ posts.
Three ways in, from a two-week diagnostic to standing technical leadership. Fixed fees, quoted after one call.
2-4 weeks
I map one AI system, or your whole estate, against the EU AI Act and ISO/IEC 42001, the standard auditors are starting to certify against. You get a risk classification, the gaps, and a prioritised roadmap with the evidence to back it.
For CTOs and compliance leads who need to answer "is this defensible?"
3-6 weeks
One workflow with a hard KPI, built with the logging, evaluation and provenance that let you prove what it does. The audit's roadmap, executed, and handed to your team with the documentation to run it.
For teams with pilots that work but cannot yet pass review.
1-2 days a week
Architecture, delivery and the conversations with your auditors and investors, as the technical lead your team can call. Infrastructure as code, evaluation harnesses and the design docs the team builds on.
For funded companies shipping AI into regulated workflows.
I'm Márton Csernai, Marci to most people (say MAR-chee). I hold a PhD in computer science, summa cum laude, and have spent fifteen years building systems where being wrong is expensive: data-centre networks, neuroscience data pipelines, a self-sovereign identity network, and now AI for regulated industries.
I work solo and hands-on. I write the architecture, run the infrastructure, ship the code, and sit in the room with your compliance team. What I don't do is policy slideware. If an AI system cannot produce the evidence an auditor asks for, I don't consider it done.
Based in Berlin, part of the year in London, working remotely with teams across the EU and the UK.
