Maroš – Python, LLM, LangGraph

Lexomat is a production AI legal-research platform for Slovak and Czech law that I founded, product-managed, and led end-to-end. It gives lawyers and legal teams an agentic research assistant, hybrid semantic/full-text search, AI document analysis and drafting, a professional directory with matching, and case-management tools — all backed by a multi-million-document corpus of legislation, court decisions, EU/CELEX sources, and ECHR rulings.

• Owned Lexomat’s product vision and roadmap from idea to production, defining the value proposition, prioritizing features, and validating them with legal professionals.
• Led end-to-end delivery of AI-powered legal research, search, document analysis and generation, case management, and lawyer directory features.
• Architected the agentic AI platform using LangGraph, ReAct agents, hybrid retrieval, and citation validation to deliver accurate, explainable legal research.
• Improved search quality with hybrid retrieval, reranking, and intelligent matching for legislation, case law, and lawyer discovery.
• Managed product delivery and engineering execution, reviewing code, mentoring a contractor, and driving critical cross-functional initiatives.
• Strengthened platform reliability through search integrity safeguards, resilient data pipelines, and production monitoring.
• Built a flexible pricing and feature-gating system integrated with Stripe and enforced through automated validation.
• Enhanced GDPR compliance with secure data retention, access controls, and legal review processes.
• Established production operations with health monitoring, LLM observability, deployment improvements, and automated quality gates.

Led the AI core of Upheal’s clinical documentation platform, responsible for the LLM systems that generate therapy progress notes used daily by thousands of clinicians. Owned the full lifecycle of prompt engineering, evaluation, RAG, agentic workflows, model selection, observability, and production quality across multiple frontier models (OpenAI, Anthropic, Google, Meta). Focused on improving accuracy, reliability, latency, and cost while enabling rapid experimentation and safe deployment of AI features.

• Owned the AI quality of Upheal’s core clinical documentation product, shipping 100+ production AI releases across 250+ note sections without quality regressions.
• Reduced LLM inference costs by ~50% while improving output quality through prompt optimization, model routing, and evaluation-driven development.
• Built an LLM evaluation platform with Langfuse, automated regression testing, LLM-as-judge evaluations, A/B testing, and release gating.
• Designed and productionized RAG, prompt orchestration, and agentic AI workflows across GPT, Claude, Gemini, and Llama.
• Evaluated and routed workloads across Vertex AI, AWS Bedrock, Azure OpenAI, and Anthropic based on quality, latency, and cost.
• Built an internal AI customer support agent using Claude Agent SDK with enterprise integrations and guardrails.
• Established AI observability and MLOps using Langfuse, Grafana, CloudWatch, BetterStack, SNS/SQS, and Superset.
• Mentored engineers on modern LLM development and advised founders on AI strategy, model selection, and cost–quality trade-offs.
• Contributed to the AI platform that supported Upheal’s $10M Series A funding by delivering production-ready, scalable LLM capabilities.

An expert-validated knowledge platform designed to organize complex domain information and support evidence-based professional decision-making. The engagement covered product definition, business-model validation, financial planning, and technical architecture for a UK innovation-funding application.

• Assessed the existing product concept and identified critical weaknesses in the business model, product scope, financial assumptions, and technical approach.
• Reframed the product around an expert-validated knowledge graph that could provide structured, traceable, and reliable information.
• Defined the target users, core workflows, product value proposition, MVP scope, and implementation roadmap.
• Redesigned the proposed technical architecture to better support scalability, expert review, and future AI functionality.
• Restructured the financial and commercial model to create a more credible path from research and development to market adoption.
• Prepared the product and technical foundations for a successful UK Smart Funding submission after approximately 12 months of unsuccessful prior work.

An enterprise climate-risk platform that assessed the financial and physical impact of climate hazards on assets and portfolios worldwide. The system combined climate, geospatial, satellite, corporate, asset, and financial datasets to quantify exposure to multiple hazards and support enterprise risk decisions.

• Built the company’s first physical climate-impact platform from the ground up, covering combined-hazard damage analysis for individual assets and global portfolios.
• Developed models that translated physical climate events into estimated asset damage and business impact.
• Shipped a financial-impact model that predicted stock-price shocks associated with adverse climate events.
• Led the development of a global wildfire-damage prediction product from research through production delivery.
• Built a company-to-subsidiary-to-asset data framework covering S&P 500 and FTSE 100 companies.
• Processed and integrated terabytes of geospatial, satellite, climate, corporate, and financial data.
• Optimized distributed data pipelines using PySpark and Databricks to improve the scalability of global analytical workloads.
• Used GeoPandas and Xarray to process large geospatial and multidimensional climate datasets.
• Transformed research models into production data products suitable for enterprise portfolio analysis.

A local food-delivery marketplace connecting restaurants, couriers, and customers across several Slovak towns (3 in 2026). The platform supported restaurant discovery, online ordering, payment coordination, delivery operations, and partner management. I co-founded the company and led the product, business, and operational execution required to bring the service from concept to a functioning marketplace.

• Co-founded the company and led the development of the business from initial concept through product launch and day-to-day operations.
• Defined the product vision, business strategy, operating model, market priorities, and growth roadmap.
• Translated business requirements into actionable tasks for software developers, designers, marketing specialists, and operational partners.
• Oversaw product development, prioritized features, reviewed progress, and coordinated delivery across technical and non-technical teams.
• Worked closely with programmers and designers to shape the customer ordering experience, restaurant workflows, courier processes, and internal operational tools.
• Managed relationships with restaurants, presented the platform’s value proposition, negotiated cooperation terms, and supported partner onboarding.
• Coordinated courier recruitment and delivery operations to maintain sufficient coverage and reliable order fulfillment.
• Secured and managed the hardware and operational equipment required to support restaurant and delivery workflows.
• Directed marketing activities, customer acquisition initiatives, local market launches, and promotional campaigns.
• Managed budgets, monitored revenues and operating costs, reviewed financial performance, and adjusted priorities to maintain commercial viability..Resolved cross-functional product, operational, financial, and partner issues as the primary decision-maker.
• Expanded the service across multiple Slovak towns and established an operating model that allowed the business to continue serving customers after the initial launch period.

A collection of production machine-learning systems for document recommendation and automated interpretation of foreign-exchange conversations. The solutions helped institutional clients discover relevant research and enabled automated extraction of trading entities from conversational messages.

• Built a history-augmented collaborative-filtering system that recommended relevant documents based on clients’ previous activity and interests.
• Productionized weekly document recommendations delivered through a client-facing chatbot.
• Improved access to a corpus of thousands of documents and reduced the time required by thousands of clients to locate relevant information. Designed and implemented an NLP entity-extraction system from the ground up for foreign-exchange conversations.
• Combined Random Forests, Gradient-Boosted Trees, Kneser-Ney n-gram language models, Naive Bayes, and Word2Vec into an interpretable classification pipeline.
• Improved the NLP architecture over time by introducing LSTM and BERT-based models using Keras and PyTorch.
• Developed more than 12 named-entity recognition models that automated chat-to-price foreign-exchange execution worldwide.
• Built scalable data and machine-learning workflows using Amazon S3, Spark, and Dask.
• Balanced model accuracy with interpretability to support production use in a regulated financial environment.