Hi, I'm Marci,
When a customer or an investor asks how your AI reached an answer, you need to show them. I build it so you can: every answer cites its source, every action leaves a trail. I work hands-on inside your team, from designing the architecture to the conversation with your compliance lead.
Previously led the research team behind Deloitte's reusable KYC credentials. 15+ years building systems where being wrong is expensive.
PhD in computer science,
summa cum laude
An auditor, a compliance lead, a customer's due diligence, your own board. I build systems that answer all four without you in the room.
Every answer carries the passage it came from, in the document your team already works from.
click through to the source
Where a decision needs a person, a named person approves it, as a step in the workflow, not a policy on a slide.
a name on every decision
Every action is recorded: who, what, when, and on which evidence.
what your auditor reads
The rules live in a file your team can read and change. A correction becomes a rule, not a rebuild.
change propagated through the system in days
Most systems stop at the first question.
"Our pilot works but would not pass review."
3-6 weeks, fixed scope
One workflow, built or hardened with the logging, evaluation and source tracking that let you prove what it does. Handed to your team with the documentation to run it.
"We need a senior lead who still builds"
Part-time or full-time, by the month
Fractional or interim CTO or architect. I own the architecture and the delivery, write code with your engineers, and take the conversations with your auditors, customers and investors.
"Would our AI hold up if someone looked closely?"
2-4 weeks, fixed scope
A review of one AI system or the whole estate: what it does, what it can prove, and where it would fail a customer's due diligence or a regulator's question. You get the gaps and a prioritised plan, mapped to the EU AI Act and ISO/IEC 42001 where they apply.
Tell me what you are building and who will use it. You get my honest read, including whether I am the right person for it.
The product reads fund documents for compliance teams, so an unsourced answer is a liability. I rebuilt how documents are split and searched, attached the source passage to every answer, and moved the agent steps into workflows where a person approves before anything is final. I also led the engineering team and the reliability work on the cloud platform (~180 managed resources), cutting service start-up from about nine minutes to under one.
TypeScript · Python · Docling · LlamaIndex · pgvector · Temporal · Terraform · Google Cloud
"We ended up with a more efficient, production-suitable RAG pipeline and higher confidence in the system's outputs."
Every answer traces to a passage in the source document.
Human approval is a step inside the workflow, not a policy.
A customer should not have to prove who they are from scratch at every institution. I led research and system architecture for KILT, an identity network where a credential is issued once and reused, from the whitepaper through the mainnet launch on Polkadot and a compliant token sale. Deloitte chose the network to issue reusable KYC and KYB credentials.
TypeScript · Rust · AWS · Kubernetes · decentralized identity · token economics
"Márton is able to work seamlessly across teams - core devs to legal to marketing and business development, and his expertise was key to the rapid growth and success of KILT in the Polkadot ecosystem."
Sources: Finextra, Deloitte on LinkedIn and Polkadot's Interview with Marton Csernai.
Every change to what the coach says took more than two weeks and a deployment. I moved the coach's rules out of code into one file the team owns and a reviewer can read, behind a four-agent backend with a safety mode for overwhelmed users. Every rule change is scored against ten test personas by an automated evaluation.
Flutter (iOS/Android) · TypeScript · multi-agent LLM backend · Promptfoo · Langfuse · OpenTelemetry · AWS
"Especially as a non-technical founder, I felt very well guided and supported."
The non-technical founder ships new features in days instead of weeks.
No lengthy app deployments to change how the AI behaves.
Ensuring GDPR-compliance by having the privacy policy verified against the code and resolving any discrepancies.
The fund is placing two deals a month with its angels, chosen from hundreds of pitch decks, and was doing it by hand, one WhatsApp message at a time. Now one partner runs a round without leaving Airtable. The model drafts the investment memo from the pitch deck and marks every gap instead of filling it and the partner approves it. The chosen angels get the deal on WhatsApp, their commitments are counted against the allocation, and when the round is covered each investor confirms and is handed to the fund's SPV platform for KYC, signing and the wire transfer. The angels never install anything or learn a new tool. The partner can answer every open question personally and the system handles the rest. The first working loop ran on WhatsApp two weeks after the first call, and the fund runs it without me.
Airtable · n8n · WhatsApp Business API · Claude API
"Very professional work. At every step Márton was patient and listened to what we needed, and we are really happy with what he delivered."
Gaps in a memo are marked, never filled by the model.
Only exact keywords change an investor's status; everything else goes to a person.
Every message to and from an investor is logged with its delivery status.
First working loop on real WhatsApp two weeks after the first call.
I hold a PhD in computer science (summa cum laude) and spent years in research before I spent years in startups. My experience spans data-centre networks, neuroscience data pipelines, identity systems, and most recently AI for regulated industries. So I work in two modes and know which one a problem needs: when to stop and think in front of a whiteboard because it saves months of work later, and when to ship by Friday.
On an engagement I write the architecture, run the infrastructure, ship the code, and sit in the room with your compliance lead. I can discuss retrieval methods with your engineers and how to present evidence with your auditor in the same meeting. What I don't do is policy slideware. If an AI system cannot produce the evidence someone will ask for, I don't consider it done.
Python · TypeScript · RAG · agents · evaluation · Temporal · Terraform · Google Cloud · Kubernetes