From a robotics lab
to the async core of
a contract-intelligence platform.

I trained as a mechatronics engineer — robotics, automation and system control at the University of Debrecen's Faculty of Engineering. Control theory teaches you one thing above all: a system without feedback drifts until it fails. Everything I've built since has been a feedback loop of one kind or another — first around actuators, now around production platforms.

I didn't wait for the degree to ship. Through the same four years I ran a Magento webshop with 130,000+ products at UNDEFASA — PHP plugin development, an automated data-cleaning system for the entire catalogue, a universal multi-carrier shipping integration, a full checkout rework. My first production system was live commerce, where every bug had a price tag attached.

At Venturi Systems I stopped being just the person who builds the system and became the person responsible for it existing at all. We set out to optimise city traffic with AI, and I architected the whole thing: a distributed simulation platform that pushed 3M+ data points through 140 simulated years of traffic on Google Cloud 96-core instances, ML-accelerated Green Wave signal optimisation, and matrix-based safety intergreen calculations so the optimiser could never trade safety for flow. Python went from a tool I used to the language I think in.

I owned the other half too. Thirty-plus meetings with municipal decision-makers, two business-to-government contracts, an external design team at EPAM managed as Product Owner, a two-person engineering team — CTO and CEO responsibility carried simultaneously. Sixteen intersections went live across three municipalities, with partnerships reaching Kyiv, Ivano-Frankivsk and Hong Kong.

Concurrent with the founder track I ran a freelance practice — 12+ client projects delivered end to end: requirements, architecture, development, deployment, ongoing optimisation. A newspaper got a cross-platform Progressive Web App with built-in subscription management and payment processing, installable on iOS and Android. E-commerce clients got a 360-degree product-photo gallery plugin, a universal multi-carrier shipping integration, complete checkout-flow redesigns.

And one client got something ahead of its time: OpenAI integrated into a live business system, automating production workflows — years before ChatGPT made that an obvious thing to want. The AI early-adopter arc that runs through everything I do now starts here.

At Pulilab I partnered directly with WHO and UNICEF project stakeholders — translating their requirements into delivered systems. A Python data-management platform for WHO. Frontend reliability for UNICEF's global user base. A 1.5-million-word content-management system delivered end to end.

The data work went deep: a Scrapy enrichment pipeline that processed 20,000+ medical clinics — images, contact data, entity extraction at scale — and a 100,000+ datapoint analytics system with multi-level aggregation from city to country to global, geolocation processing and statistical outlier detection. On top of it, an OpenAI-powered content-strategy system with automated A/B testing, built in 2021–22 — still pre-ChatGPT. I owned the whole deployment lifecycle: Docker, CI/CD, Django REST APIs on PostgreSQL, Nuxt.js in front, Jupyter and Pandas for the reporting.

Affinitext is a contract-intelligence platform that governmental, defence and infrastructure clients run their post-execution obligations on — multi-billion-dollar programs where a missed clause-driven deadline can trigger eight-figure liquidated damages. I work the hardest subsystems: task orchestration, identity and SSO, document linking and clause graphs, cloud infrastructure on GCP, and reliability for long-running data pipelines.

The defining work: sole author of the platform's async substrate — a managed Celery framework, +10,787 lines across 42 files, multi-tenant context restoration, 50+ legacy jobs ported off a jobserver that used to crash without logs. Two years later I'm still its runtime owner. When production locks up, I command the incident. When the user-log report hung for 15.4 seconds, I took it to 0.67 at the query layer. When the core checkout-status workflow needed rebuilding, I carried it through ten months and eight versions until status finally meant what it said.

Unprompted, I went looking for the failure modes nobody had reported: four pre-disclosure vulnerability reports in thirty days. A dangling-IP subdomain takeover caught through DNS archaeology. CSRF tokens issued from a Mersenne-Twister PRNG whose internal state — and therefore every future token — was recoverable. An unauthenticated callback that could disable a target's SMS multi-factor authentication. A leaked-credential audit across public breach data. Each written in executive register, framed against ISO 27001, with a remediation plan attached.

In parallel I built the team's AI-engineering practice: a Claude Code slash-command suite in daily production use, an MCP-driven multi-source evidence pipeline, and a quantitative analysis of a 1,000-merge-request review corpus that grounds a tooling proposal automating roughly a sixth of recurring review feedback. The lab runs hot alongside it — concurrent Gemini agents scoring trading hypotheses over a Celery + FastAPI backtesting engine, native Swift apps, an MCP server published on PyPI.