Skip to content
Maroš Jančo

Senior AI engineer for production LLM systems.

I help founders and engineering teams ship agentic, RAG and LLM systems that actually work in production — evaluations, context engineering, MLOps, full-stack.

Currently taking a small number of consulting engagements. Based in Slovakia, working with clients across Europe and the UK.

Currently available for new engagements.

See recent work ↓Email me →

Worked with and at:

Selected Work

Upheal · Senior AI Engineer · Nov 2023 – Present

How do you ship LLM products that don't regress in production, across 100+ releases and multiple model generations?

Owned prompts, models and quality for AI-generated clinical progress notes — the core product of a Best-Startup-Award-winning documentation platform for therapists.

  • Quality mattered across 250+ documentation sections: dozens of prompt flows, RAG where retrieval helped, agentic patterns where multi-step reasoning was needed, and a family of AI text-editing tools for clinicians.
  • Shipped 100+ production releases without quality regression across multiple model deprecations and new arrivals (Gemini, Claude, GPT-*, Llama).
  • Cut AI cost ~50% with no quality regression — backed by an LLM-as-judge evaluation framework on Langfuse (datasets, eval runs, trace-level flow debugging) and an A/B-testing pipeline that gated releases, caught regressions and shortened prompt iteration.
  • Benchmarked Vertex AI, AWS Bedrock, Azure OpenAI and Anthropic continuously for quality, latency and cost; routed traffic by use case rather than committing to one provider.
  • Automated a Claude-Agent-SDK customer-support agent (internal-tool integration, hardened prompts, guardrails). Ran LLM observability and MLOps; reported AI roadmap and cost trade-offs directly to founders, and lifted team velocity — Claude Code in daily workflow, CI/CD, internal RAG/prompting/agentic sessions.

Python · TypeScript · Langfuse · Vertex AI · Bedrock · Anthropic API · Grafana · BetterStack · CloudWatch · SNS · SQS · Superset

Cervest · Senior Data Scientist · Jun 2021 – Jun 2023

How do you ship production ML over TBs of multi-source data, where the predictions bear real financial risk?

Built from scratch the first physical-climate-impact platform assessing combined-hazard damages on assets and portfolios worldwide. The company won the Global Impact 50 Award 2023.

Led a global wildfire-damage prediction product end-to-end. Built a company → subsidiary → asset framework for SP500 and FTSE100 firms; shipped a financial-impact model predicting stock-price shocks from adverse climate events. Optimised distributed pipelines over TBs of geospatial, satellite and financial data.

PySpark · Databricks · Geopandas · Xarray · Kubernetes · Python

BNP Paribas · Machine Learning Researcher · Jun 2016 – Jun 2021

How do you ship NLP models that run daily in mission-critical workflows, year after year?

Built and deployed NLP models for a global investment bank's AI lab.

Built a history-augmented collaborative-filtering recommender for client document recommendations; weekly chatbot recommendations save thousands of clients hours of search across thousands of documents. Designed and shipped an NLP entity-extraction system combining Random Forests, GBT, Kneser-Ney n-gram LMs, Naive Bayes and Word2Vec into one classifier; later scaled with LSTM- and BERT-based architectures. A dozen+ resulting NER models now automate chat-to-price FX execution worldwide on a daily basis.

Python · PyTorch · Keras · Spark · Word2Vec · BERT

Lexomat · Founder · 2024 – Present · lexomat.sk →

How do you ship LLM systems that domain experts will trust to use daily?

Solo-founded AI legal-research chat for Slovak and Czech law, indexing millions of legal documents. Used by legal professionals through FinAI s.r.o.

Drove product, design, agentic AI flows, GDPR compliance, pricing and infrastructure end-to-end (one contractor on delivery). Built the citation-resolution layer for ambiguous pre-1993 Czechoslovak legal references and a backend-resolved URL architecture for slov-lex.sk to eliminate LLM URL hallucination.

Next.js · TypeScript · Python · FastAPI · Supabase · PGroonga · Vertex AI

Open Source

Souli · open-source · github.com/dzino-app/souli.app →

How do you build a useful AI companion without sending plaintext chat to a third-party model?

End-to-end encrypted AI companion that gamifies personal growth across social, health, career and personal aspects. Privacy-first architecture — chat content is encrypted end-to-end, including in transit to and from the model.

Open-sourced May 2026. Full-stack TypeScript + privacy engineering.

How I Work

Engagements

Most engagements start with a conversation like: “our RAG retrieves the wrong documents and we can’t diagnose why,” “our LLM costs are 4x what we modeled,” “our agent demos beautifully but breaks with real users,” or “we need an evaluation framework before our next release.”

Typically 1–3 months, scoped per problem. I take on a small number of clients at a time.

Pricing

Engagements priced per project or per day, scoped in the first call. I’ll send a clear range within 24 hours of our initial email exchange.

Working style

Remote-first from Slovakia with regular travel — comfortable working hybrid across major EU hubs (London, Berlin, Munich, Amsterdam, Zurich, Prague, Vienna) on a roughly monthly cadence. Async-friendly, English, EU/UK timezones.

About

Maroš Jančo

I’m a senior AI engineer with 9+ years shipping production ML, NLP and LLM systems. I trained at Imperial College London (BSc Mathematics with Statistics for Finance, First Class) and UCL (MSc Machine Learning), then spent four years at BNP Paribas’s AI lab, two years at Cervest in climate-risk modelling, and the last two and a half years leading AI engineering at Upheal.

Alongside employment I’ve founded two AI products: Lexomat, a legal-research chat used by lawyers across Slovakia and Czechia, and Souli, an open-source end-to-end-encrypted AI companion. Earlier, in 2019, I co-founded PayToEat, a food-delivery app still used by customers today. The combination of senior IC engineering, founder experience, and direct customer-facing product work is what I bring to engagements.

I’m a generalist by preference — I want to own a problem from prompt to production, not specialise into one layer of the stack. The kind of work I do best is the messy middle: making a demo-grade AI system into something that ships, scales, costs less than it earns, and doesn’t break in surprising ways.

Outside work I read daily about investing, economics, agentic systems, and AI advancements. I speak English, Slovak, Czech and beginner German.

Writing

Coming soon: notes on building production LLM systems, AI evaluations, and lessons from shipping Lexomat and Souli.

All writing →

Get in touch

If you’re working on a production LLM system and want a senior pair of eyes — or you’d like to discuss a longer engagement — email me a paragraph about what you’re working on.

maros@marosjanco.com

I read every email personally and reply within 48 hours. If we’re a fit, the next step is a 30-minute call.