AI consultant at J&T Banka in Prague. My job is taking generative AI and agents past the risk committee and into live operation: real accounts, audit trails, people in the loop, and models that run on your own hardware when the data can't leave the building. Available for consulting across the EU, in Czech or English.
Production, in my vocabulary: approved by risk, wired into core systems, logged for audit, supervised by people, and running every working day.
Worth itChoosing the work
Not every process wants a model. I look for the work that is high-volume, well documented, and expensive when it goes wrong: document processing, back-office workflows, the operational plumbing nobody demos. Then build versus buy, and a sequence of adoption that survives contact with compliance, security, and the org chart.
GuardrailsDesigned before the demo
Agentic systems earn permission to act through test suites, human-in-the-loop oversight, and audit trails that a reviewer can actually read. Transparent and explainable by design, because in finance a black box is a liability, not a feature.
Wired inConnected to core systems
A model that reads exports is a report. A model wired into events, APIs, and operational data can act on a real account, with the same controls as any other system in the bank. That integration work is where most AI projects quietly die, and where I spend most of my time.
LocalOn your own hardware
Regulated data often cannot go to a cloud API, so I run models where the data lives. My public benchmark: the unpruned 2.78-trillion-parameter Kimi-K3, served across two Mac Studios over Thunderbolt RDMA with MLX, at 24 tokens a second aggregate under 32 concurrent requests. The converter, the expert-pruning tooling, and the reproducible records are open source on GitHub.
LiveShipped, then kept shipped
Production means monitoring, fallbacks, and a person who can stop it. Move fast, with safeguards that match the stakes. The bar is simple: it has to survive the risk committee, and then it has to survive a Monday morning at quarter end.
End to endHead of Digital Experience, 2019–2025
Before AI, I owned J&T Banka's digital product end to end as Head of Digital Experience (2019–2025): the next-generation mobile bank, live since 2024, with banking, investments, and trading in one app and response times cut from 350 ms to 80 ms, on an event-driven platform that carries 2M+ messages a day across 20+ core systems, delivered in seven and a half months. Same regulator, same constraints, same standards.
The digital banking work is publicly documented in the Trask case study, the Ackee case study, and the Krásní lidé, krásné appky podcast episode (May 2025).
For banks, insurers, and other regulated enterprises. In Prague or remote, in Czech or English.
The future of banking isn't about replacing humans with AI. It's about augmenting human capabilities with intelligent systems that handle complexity at scale.
The wins came from aligned incentives, iterative delivery, and a relentless focus on measurable outcomes. I learned as much from the failures as from the successes, and I will tell you about both.
Project lead, CTO, product manager, head of digital, and now AI. Same job, in one sentence: bridge what technology can do and what a regulated institution can trust.
Shipping small things end to end keeps me honest about what "production" costs.
A free-to-play portrait co-op merge tower defense for Android and iOS. Merge defenders in your lane, share one health pool, and play solo with an AI ally or online with another player.
An MLX port and distributed runner for Moonshot's 2.78T-parameter multimodal MoE: a streaming converter for a 5.6 TB source, REAP expert pruning, per-language expert-overlap analysis, and reproducible two-machine benchmarks over Thunderbolt RDMA.
Privacy-first stress monitoring for iPhone and Apple Watch. Fully on-device, no servers, no analytics, with insights from Apple's on-device foundation models.
A gentle candy block puzzle for kids and families on iPhone, iPad, and Mac. No ads, no purchases, works offline.
An AI consultant at J&T Banka in Prague and the bank's former Head of Digital Experience, specializing in agentic AI systems, enterprise AI integration, and intelligent automation in regulated financial services. Czech spelling: Michal Komár.
Take generative AI from slideware to production: pick the use cases that matter, design the guardrails and human oversight, integrate with core systems, and ship inside real compliance, security, and audit constraints.
Yes. I'm based in Prague and work across the European Union and globally, remote-friendly, in Czech or English.
LinkedIn is fastest: linkedin.com/in/michalkomar. I'm also on X and GitHub.
Putting AI into a business that answers to a regulator? That is the conversation I want to have.
Bring the hard version: agents touching real accounts, models next to core systems, an organization that has seen three transformations already. I answer on LinkedIn within a day or two.
Jsem Michal Komár, AI konzultant v J&T Bance v Praze, dříve Head of Digital Experience. Uvádím generativní AI do produkce v regulovaném prostředí: agentní systémy, integrace do podnikových systémů a inteligentní automatizace. Předtím jsem jako Head of Digital Experience vedl celý digitální produkt banky včetně nové mobilní banky, která je v provozu od roku 2024. Konzultace probíhají česky i anglicky, v Praze i vzdáleně. Nejrychlejší kontakt je LinkedIn.