MARKO RUMAN

STATUS: FINISHING PHD (CAS/UTIA) · ACCEPTING PROJECTS

// SPEC 01 : APPLIED MACHINE LEARNING & FULL-STACK PRODUCT ENGINEERING

I build machine-learning systems and the products around them.

From statistical models and sparse-data recommenders to full-stack AI applications—I turn mathematical rigor into deployed software that people actually use.

I’m Marko Ruman. I combine an academic research background in AI, decision support, and statistics (PhD thesis submitted) with full-stack software development. I don’t just train weights or write LLM prompts; I engineer the entire system: from framing the business decision and mastering imperfect data to architecting backends, databases, and responsive interfaces.

// TELEMETRY & CREDENTIALS

LIVE_SPEC

ACADEMIC ANCHOR

PhD Thesis Submitted

AI/ML Research · CAS & UTIA context

NEXT CHAPTER FOCUS

Full-Time Practical Builds

Applied ML, Consulting & Product Engineering

CORE DIFFERENTIATOR

Small Data & Offline Systems

Extracting high-value signal where massive deep learning cannot run

STACK OWNERSHIP

End-to-End Delivery

Svelte · Python · PHP · PostgreSQL · Expo · MCP

COLLABORATION: REMOTE / GLOBAL

●

// 00 — THE ARCHITECTURAL DIVIDE

Good models are only part of the solution.

“A model without software around it is an isolated benchmark. A product without rigorous statistical framing is a fragile toy.”

01 / NOT JUST ACCURACY METRICS

Decision-Support Over Isolated Predictions

A 98% accuracy score is meaningless if the prediction doesn’t empower an operator to take an action, or if it breaks when operational inputs change. I frame problems around the business decision, user workflow, latency budget, and verifiable payoff.

02 / STRUCTURED STATE OVER RAW CHAT

Evidence, Provenance & Auditable Schemas

Chatbots that generate unstructured summaries drift into hallucination and lose source provenance. I build structured state layers that extract schema-driven claims, attach immutable evidence links, and model confidence for downstream software.

03 / SMALL & OFFLINE DATA REALITY

Extracting Value Without Big-Tech Budgets

Most organizations do not possess billions of clean web interaction events. I specialize in extracting practical predictive power and personalized recommendations from sparse, messy, and physical offline retail transaction logs.

// 01 — SELECTED WORK & APPLIED SYSTEMS

Products & Architectures I’m Building

A transparent showcase of self-directed products, open architectures, and algorithmic prototypes. Rigorous end-to-end engineering, zero invented metrics.

01 // FLAGSHIP PRODUCT

LIVE ALPHA · ELEVIST.AI

ROLE: FOUNDER & BUILDER

Elevist — An AI Execution System for High Performers

“Connecting long-term goals with daily reality through an automated AI reflection feedback loop.”

THE PROBLEM

High performers track goals in one tool, tasks in another, habits in a third, and daily notes in random docs. Even with a plan, the feedback loop between intended priorities and real execution is broken. Context decays, execution drifts, and retrospectives are forgotten.

THE ENGINEERING & AI HARNESS

Elevist unifies goals, metrics, tasks, habits, and natural language check-ins. Rather than a generic chatbot, AI is applied to a structured Friday Review workflow: comparing what was planned against what actually occurred, computing consistency scores, identifying high-leverage bottlenecks, and resolving the “unstuck” state.

Plan-vs-Reality AI Review

Leverage Tagging (000/h vs 0/h)

Consistency Score Algorithm

Unstuck Diagnostic

VISIT ELEVIST.AI ↗READ CASE STUDY →

SYSTEM ARCHITECTURE

Frontend Application

Svelte · Responsive Reactive UI

Backend & Persistence

PHP Backend · PostgreSQL State Engine

AI Context Pipeline

Structured Context Synthesis & Evaluation

Deployment Infrastructure

Coolify Managed Containers · Linux

WHAT THIS DEMONSTRATES:

Applying AI to concrete human workflows rather than generic conversational prompts. End-to-end full-stack ownership.

02 // STATE & CONTEXT ENGINE

OPEN ARCHITECTURE · GITHUB

ROLE: ARCHITECT & CREATOR

ClueState — Evidence-Backed User & Customer State Engine

“Turning unstructured interactions into structured, evolving state—with verified provenance behind every claim.”

THE PROBLEM

Applications extract insights from conversations, notes, and activity. But raw text summaries lose provenance, become obsolete, hide internal contradictions, and cannot be consumed by deterministic SQL databases or rule-based backends.

THE ARCHITECTURAL SOLUTION

ClueState functions as a connector and converter between unstructured inputs and a predefined state schema. LLMs extract candidate claims, while the engine permanently links each claim to its source evidence, assigns confidence scores, and explicitly models changing or superseded facts.

Predefined State Schema

Immutable Provenance Links

Confidence & Decay Scoring

Contradiction Reconciliation

INSPECT ON GITHUB [ GITHUB.COM/CLUESTATE ] ↗READ ARCHITECTURE STUDY →

POTENTIAL APPLICATION DOMAINS

Deep Product Personalization

Persistent, auditable user preferences that adapt over time without drift.

Retail & Loyalty Customer Profiles

Synthesizing customer behavior and interaction notes into queryable CRM attributes.

Audit-Proof Customer Support

AI features that cite exact source evidence before triggering customer-facing actions.

WHAT THIS DEMONSTRATES:

A systems-level view of LLM integration. Replacing naive prompt dumps with rigorous data modelling and provenance guarantees.

03 // FULL-STACK PLATFORM

CROSS-PLATFORM RUNTIME

ROLE: BUILDER & ENGINEER

CoevoApps — Single-Project Web & Mobile Platform

“A more direct path from an idea to production software, without compromising on code ownership.”

THE PROBLEM

Building cross-platform products requires juggling disconnected stacks: web frontend, mobile codebase, backend APIs, database migrations, and deployment. No-code platforms promise speed but trap companies in proprietary walled gardens.

THE COMPILER & MCP INTEGRATION

CoevoApps provides a single project model that compiles clean code for both web (Svelte) and native mobile (Expo / React Native). It incorporates direct PostgreSQL management, Coolify deployment, and a full Model Context Protocol (MCP) server that empowers AI agents to inspect, edit, and build applications programmatically.

Shared Web & Mobile Compiler

PostgreSQL & SMTP Runtime

Coolify 1-Click Publishing

Model Context Protocol (MCP) Server

100% USER OWNERSHIP OF GENERATED SOURCE CODE

STACK & INFRASTRUCTURE

Target Platforms

Web (Svelte) · Mobile (Expo / React Native)

Backend & DB Layer

PHP Compiler · Managed PostgreSQL

Agentic Tooling

Full MCP Tool Schema & Automation

WHAT THIS DEMONSTRATES:

Product architecture at developer-tooling scale. Unifying developer velocity, visual design, and software infrastructure.

04 // APPLIED ALGORITHMS

CAPABILITY DEMO & PILOT FRAMEWORK

DOMAIN: RETAIL & RECOMMENDER SYSTEMS

Retail Personalization for Small, Sparse & Offline Data

“Delivering measurable recommendation uplift with the data businesses actually have—not hypothetical tech giants’ data lakes.”

THE COMMERCIAL OPPORTUNITY

Brick-and-mortar retail and loyalty programs accumulate transaction logs and purchase receipts, but lack real-time digital clickstreams or massive data engineering departments. Standard deep collaborative-filtering architectures fall over on sparse histories and extreme cold-start shopper distributions.

THE ALGORITHMIC APPROACH

A hybrid ranking system that marries basket affinity matrices, inter-purchase temporal decay, category affinity vectors, and business rules. Designed to run with zero GPU overhead, low latency at POS or mobile app checkouts, and clean offline evaluation frameworks.

Sparse Basket Association Rules

Cold-Start Customer Profiles

Inter-Visit Timing Dynamics

Zero-GPU Computational Footprint

TARGET BUSINESS OUTCOMES

Personalized Cross-Sell Offers

Targeted receipts and app notifications based on basket complement probability.

Customer Lapsed-Visit Prevention

Statistical detection of abnormal gaps in repeat store visits.

Margin Protection

Discount only when required, avoiding blanketing high-intent buyers with margin-eroding promotions.

WHAT THIS DEMONSTRATES:

Applying pragmatic mathematical intelligence to real, imperfect offline data rather than chasing theoretical academic hype.

// 02 — PRACTICAL COMMERCIAL CAPABILITIES

What We Can Build Together

Framed as concrete problem-solving deliverables, not an endless menu of disconnected technologies. Each service defines the business challenge, the technical work, and the concrete outcome.

CAPABILITY 01

Applied Machine Learning & Statistical Modelling

Ideal For:

Companies with existing operational data that need predictive classification, scoring, demand forecasting, or decision-support tools.

Problem framing, rigorous exploratory data analysis, establishing transparent baseline benchmarks, training predictive models where justified, and integrating model inferences directly into backend application workflows.

TYPICAL DELIVERABLES:

Feasibility & data audit, reproducible benchmark code, evaluated production-grade model, and inference API integration.

CAPABILITY 02

Recommendation Systems & Personalization

Ideal For:

Retailers, marketplaces, loyalty programmes, and SaaS products wanting higher relevance with limited or offline data.

Product/content ranking, customer behavioral segmentation, cold-start handling, and hybrid personalization combining business rules, transaction history, and statistical affinity models.

TYPICAL DELIVERABLES:

Recommendation strategy, offline evaluation suite, ranking engine service, and database schema integration.

CAPABILITY 03

LLM Applications, Structured Memory & State

Ideal For:

Applications that collect unstructured text (notes, calls, transcripts) and need reliable, auditable, queryable state.

Transforming messy conversational data into strict schema attributes, attaching immutable source evidence to claims, managing confidence scoring, and preventing hallucination drift across application sessions.

TYPICAL DELIVERABLES:

State schema definition, extraction & update pipeline, evidence/provenance graph, and working application connectors.

CAPABILITY 04

Full-Stack AI Product Development

Ideal For:

Founders and product leaders who have a concept or ML capability and need the complete software built around it.

Technical architecture, reactive frontend interfaces (Svelte), mobile workflows (Expo/React Native), robust backend infrastructure, PostgreSQL schemas, and full containerized deployment without handing off to third-party contractors.

TYPICAL DELIVERABLES:

Complete production application, source repository, database migration scripts, and deployed infrastructure.

CAPABILITY 05

Technical Discovery & PoC Prototyping

Ideal For:

Teams that want to de-risk a high-stakes AI assumption before committing to a costly, lengthy build.

A 2-to-3 week time-boxed sprint: evaluating data feasibility, establishing if statistical baselines suffice, designing minimal system architecture, and testing the riskiest assumption with a working proof-of-concept.

TYPICAL DELIVERABLES:

Feasibility blueprint, interactive PoC prototype, and an honest build-vs-buy roadmap.

// ENGAGEMENT FORMATS

How We Can Collaborate

[ 01 ]

Discovery Sprint

Time-boxed assessment of your data readiness, algorithmic feasibility, and architecture recommendations.

[ 02 ]

Proof of Concept

Rapid, high-conviction functional prototype focused on validating the riskiest technical hypothesis.

[ 03 ]

End-to-End Build

Complete ownership from data pipeline to model training, backend API, and frontend user interface.

[ 04 ]

Senior Advisory

Part-time or project-based ML engineering leadership embedding with your existing technical team.

// 03 — ENGINEERING PROTOCOL

From Problem to Production Software

“A disciplined 6-stage lifecycle designed to eliminate guesswork, avoid premature over-engineering, and deliver verifiable software.”

STEP 01 // FRAMEWORK

Understand the Decision

Clarify the human decision, system workflow, or business constraint that needs improvement. Define concrete metrics for success before writing a single line of model code.

STEP 02 // REALITY

Inspect Data Reality

Audit the real data: sparsity, collection artifacts, sampling biases, and existing system latency. Identify what can realistically be learned from available evidence.

STEP 03 // SELECTION

Simplest Viable Intelligence

Establish a strict baseline. Choose the simplest approach that solves the problem reliably: statistical rules, classical ML, recommendation heuristics, or LLM extraction.

STEP 04 // EVALUATION

Build & Benchmark

Implement focused prototypes. Evaluate rigorously against baseline performance, error distribution, and operational boundary failure conditions.

STEP 05 // INTEGRATION

Full-Stack Harness

Embed the capability into real software: APIs, database schemas, background workers, security constraints, and intuitive user interfaces.

STEP 06 // ITERATION

Operational Feedback

Observe how the system operates under actual human use. Monitor drift, log edge cases, and iterate where improvements produce tangible economic value.

// 04 — BACKGROUND & PHILOSOPHY

Research Mindset. Product Engineering.

“I believe the most valuable engineers are those who understand deep statistical theory, but refuse to stop before the code is running in users' hands.”

I’m Marko Ruman, a machine-learning engineer and full-stack product developer. I am finishing my PhD in AI/ML (thesis submitted), with academic research connected to the Czech Academy of Sciences and the UTIA context.

Rather than pursuing a conventional academic career path, my next chapter is entirely dedicated to practical, high-impact commercial projects: building products, advising technical founders, and deploying robust ML and full-stack systems that solve actual business bottlenecks.

I approach engineering with intellectual honesty: I do not invent vanity numbers or pretend every problem requires a billion-parameter model. I design systems that can be rigorously measured, safely maintained, and integrated into living production codebases.

EDITORIAL POLICY & VERIFICATION

This portfolio represents real systems, verified code, and clear product concepts. Academic citations, publication DOIs, and detailed CV credentials are provided directly upon inquiry.

// PRODUCTION TECHNICAL STACK

CORE LANGUAGES

Python

PHP

TypeScript / JS

SQL (PostgreSQL)

Bash / Shell

AI & MACHINE LEARNING

Statistical Modelling

Recommender Engines

Decision-Support Theory

LLM Schema Extraction

Small-Data Statistics

FRONTEND & MOBILE

Svelte

Expo / React Native

Modern Responsive CSS

Lexicon / AST Runtimes

INFRASTRUCTURE & AGENT PROTOCOLS

Model Context Protocol (MCP)

Docker & Linux

Coolify Orchestration

Git / GitHub CI

// 05 — INITIATE CONVERSATION

Have a problem worth solving?

“Tell me what you are trying to build, what data or systems you currently have, and where your team is getting stuck.”

We can work out whether applied machine learning, a lightweight statistical model, an LLM structured state pipeline, or a full-stack product sprint makes the most commercial sense.

LOCATION:

Prague, Czech Republic // Global Remote

CURRENT ENGAGEMENT CAPACITY: OPEN FOR Q1/Q2 DISCUSSIONS

PROJECT INTAKE TERMINAL

YOUR NAME / COMPANY

WORK EMAIL

PRIMARY ENGAGEMENT INTEREST

PROBLEM OR SYSTEM BRIEF

PREFER DIRECT EMAIL? REACH OUT AT

MARKO.RUMAN@GMAIL.COM