Luka Abramovic

Luka Abramovic is a Croatian entrepreneur and founder and CEO of Adamant Code, a Rijeka-based company specializing in custom software, business automation, and applied AI systems. He leads the company’s product direction and works with businesses to understand operational problems, identify where technology can create meaningful value, and translate those requirements into software products and systems delivered by Adamant Code.

His work spans product strategy, solution design, and delivery across document intelligence, internal operations, customer support, AI coaching, recommendation systems, voice interfaces, marketplaces, and other custom business applications.

Early life and education

Luka Abramovic was born in Rijeka, Croatia, in 1994. He studied marine engineering and completed his bachelor’s degree in 2017 before beginning a career at sea.

While working aboard ships, he developed an interest in software and began studying programming independently, often with limited internet access. What began as a personal interest eventually became a career transition, leading him from marine engineering into professional software development and, later, entrepreneurship.

Marine engineering and career change

After graduating, Abramovic worked at sea as an engineering cadet while preparing for engineering-officer certification. During his time aboard ships, he continued studying software development and building projects in his free time.

After approximately two and a half years in the maritime industry, he decided to change careers. Following his final tour at sea, he moved into professional software development, where he gained experience building applications and working with clients before eventually founding Adamant Code.

Technology career and founding Adamant Code

Before founding Adamant Code, Abramovic worked professionally in software development, giving him a technical foundation in building applications and working directly with clients. He founded Adamant Code in 2023 and subsequently moved from hands-on development toward leading product strategy, solution design, client relationships, and delivery.

Adamant Code is based in Rijeka and works with clients in the United States, United Kingdom, and European Union. The company describes itself as a product and technical partner rather than solely an implementation contractor. Its services include process automation, systems integration, custom applications, improvements and rebuilds of existing software, product design, full-stack development, AI engineering, and continuing maintenance.[1]

The firm’s process begins with the people, business rules, existing systems, and constraints behind a problem. It then determines whether the appropriate intervention is an integration, an improvement to an existing application, a new software product, or an AI system. This approach reflects Luka’s view that the availability of generative AI does not by itself justify using it in every workflow.[1][2]

By 2026, Adamant Code reported more than 20 completed projects and presented a three-person core leadership and delivery team in Rijeka. Its website identified Luka as CEO, Matej Paus as chief technology officer, and Matej Poljanic as head of delivery.

Document intelligence and internal operations

One of Adamant Code’s documented enterprise systems is the TCE Document Intelligence System. TCE engineers were required to inspect technical bid packages that could contain between 5 and 20 PDF files and approximately 5,000 to 50,000 pages. The previous process relied on opening files separately, using basic text search, reviewing local copies, or consulting printed and highlighted pages.[1]

Adamant Code built a natural-language search product around the requirement that engineers be able to verify every generated answer. Responses therefore include citations, page references, exact supporting paragraphs, and highlighted source passages. The wider application also manages documents, projects, administrators, teams, users, and access permissions.[1]

The system entered production and continued to be used by TCE. Based on interviews and client estimates rather than controlled usage logs, daily document-search time declined from approximately 60 minutes to approximately 10 minutes per engineer. The result represented an estimated 83% reduction while keeping the underlying evidence visible for human review.[1]

Adamant Code also developed a confidential internal AI operations platform during a two-year client relationship. The platform brought recurring data pipelines, analytics, dashboards, natural-language requests, AI-assisted generation, permissions, quality review, and monitoring into a shared product. It served 44 internal users, connected more than 400 properties, and processed an average of about 220,000 rows per day.[1]

Customer operations and conversational systems

For Ergo Global, Luka’s team developed an AI customer-operations system spanning website chat, internal application chat, email inboxes, approved knowledge resources, self-assessment, scheduled follow-up, and human handover. The project treated each incoming conversation as an operational routing decision rather than merely a request for generated text.[1]

Routine questions could be answered using approved manuals, website information, and relevant order data. Cases involving pricing, bugs, dissatisfied users, sales opportunities, unknown answers, or direct requests for a person were escalated. This created a defined boundary between automation and work requiring contextual or sensitive human judgment.[1]

Over five months, the platform managed 11,782 conversations across seven support inboxes and exchanged 15,120 messages. Website chat achieved a typical first-response time of 36 seconds. An associated follow-up system sent 3,019 personalized check-ins across 1,749 employee profiles, an amount modeled as equivalent to roughly 150 to 250 hours of manual outreach.[1]

Leelou AI Coach

Luka was involved in translating TNM Coaching’s approximately 25 years of coaching experience and proprietary methodology into Leelou AI Coach. Rather than designing a general advice chatbot, the teams sought to create a system that used coaching questions to guide users toward their own reflection and decisions.[1][3]

Adamant Code learned TNM’s techniques from its coaches, converted those techniques into product and behavioral requirements, and repeatedly tested the system with domain specialists. Leelou supports structured sessions, text and voice interaction, multiple languages, continuity between sessions, accountability, and follow-up by email or text message. Organizations receive aggregate engagement information without access to private coaching conversations, while safeguards address discussions that move outside the intended coaching context.[1]

Among onboarded regular users, 95 percent completed at least one session, average package completion was 45 percent, and recorded analytics showed an average session length of 13 minutes.[1]

In January 2026, Luka appeared with Zoran Todorovic on the TNM Unplugged episode “Building Believable AI Coaches with Luka Abramovic.” The discussion addressed the product’s architecture, data structures, behavioral design, ethical boundaries, crisis protocols, and the relationship between human and AI coaching. Luka presented the project as an effort to enhance access to coaching and human capability rather than to impersonate or replace human coaches.[3]

Marketplaces, voice interaction, and recommendations

Adamant Code built the GTS Innovative AI-Assisted Marketplace from a founder’s initial concept and rough wireframes. Luka’s team defined the product’s user journeys, permissions, account types, pricing logic, edge cases, and AI-assisted invention workflow. The completed platform contained more than 50 screens, over 15 core workflows, and a six-step process for developing and presenting an invention. It supported inventors, service providers, investors, and administrators through discovery feeds, profiles, messaging, jobs, reviews, notifications, guided onboarding, and an NDA-based gate for viewing protected invention information.[1]

For Juggle Ease, the company added a multilingual real-time voice assistant to an existing family-productivity application. The feature connected natural speech to real changes in lists, tasks, reminders, trackers, and events through the application’s existing API. The scope included speech handling, intent recognition, action mapping, confirmation by voice, and mobile audio behavior across Android and iOS. It was delivered after approximately two months and supported five categories of product actions. A demonstration showed the assistant creating and updating a list in English before continuing the interaction in Croatian.[1]

The Spinio AI Recommendation Platform combined a web and mobile interface with preference logic, structured catalog retrieval, a knowledge layer, and external service connections. Its catalog covered approximately 20,000 items from about 1,000 providers. Adamant Code trained a custom text-to-SQL engine so that exact requests involving catalog properties could be translated into structured database queries rather than handled through imprecise semantic similarity alone.[1]

The platform allowed users to begin with topic shortcuts or free-text questions, receive explorable result cards, continue earlier conversations, and access comparable experiences on desktop and mobile. Preference signals and entity tags contributed to ranking, while controlled knowledge sources supported explanatory questions beyond catalog filtering.[1]

Writing and views on applied AI

Luka has published writing about computer-science education, learning software development, AI architecture, entrepreneurship, and professional freelancing. His autobiographical essay “From 10 MB/Day to a 6-Figure AI Agency” described his progression from ship engine rooms and restricted internet access to professional software work and the creation of Adamant Code.

In “When RAG Struggled and SQL Provided the Answer,” he examined a natural-language search system for structured records. He argued that retrieval-augmented generation is useful for unstructured documentation but unreliable for exact filtering involving names, numerical values, payment methods, or ranges. The resulting architecture routed structured requests to a trained text-to-SQL system and knowledge questions to a retrieval-based system.

On May 1, 2026, Fiverr Community published Luka’s essay “What makes a great AI freelancer? The same things that always did.” He identified judgment, communication, and a focus on value as enduring professional skills. He argued that AI should not be inserted into deterministic processes that require the same exact result every time and is better suited to work where variation is useful or acceptable.[2]

He also emphasized the importance of explaining complex AI workflows to clients. Operations involving multiple model calls, database lookups, and processing stages can create latency and cost that are not visible from the interface. In his account, communicating those constraints, showing progress, and setting accurate expectations remain central to effective delivery.[2]

Across Luka’s projects and writing, a consistent principle is that AI is a component within a larger product rather than a substitute for product design. Adamant Code’s systems combine models with source evidence, structured data, permissions, integrations, privacy rules, review processes, escalation paths, and domain-expert testing. Luka’s public position is that useful AI depends less on novelty than on selecting an appropriate problem, understanding the operating environment, and delivering a system that produces measurable value.[1][2][3]