trilar~/ai-reliability — Jure Trilar, home
Available for select strategic advisory roles, expert evaluation, and keynotes

Frontier AI that works today,
and still works tomorrow.

I help companies and public institutions take LLM and agentic systems from demo to dependable: the right use cases, measured reliability, and designs ready for the next models and the next EU rules.

[Initiate Collaboration] Read the notes

Dr. Jure Trilar · Researcher from Slovenia
University of Ljubljana · Faculty of Electrical Engineering

$ npx reliability-audit --system support-agent --reruns 32→ 412 eval cases · 4 dimensions · 3 fault typesconsistency0.94calibration0.81robustness0.88recoverability0.76✓ evidence → ai-act/accuracy-robustness.md
Illustrative output of a reliability audit: four dimensions, one threshold breach, an evidence file.

01 Point of view

Everyone has access to the same models. The advantage is making them work together dependably.

Powerful AI no longer sets organisations apart. What does is how quickly — and how safely — they turn it into operations people can rely on. That is the gap I work in.

Pragmatic futurism

I follow where AI research and standards are heading over the next 12–24 months, and turn it into decisions you can act on this quarter.

Impact, not just AI

We agree on what success means before we start: measured reliability plus one business outcome. “We implemented AI” doesn’t count.

Tomorrowing today

Everything I deliver works now, and is built to survive the next model, the next scale-up and the next regulation.

02 Research

Reliability of AI systems, measured in multiple dimensions.

My current research is a systematic approach of how reliability is measured in LLM-and-Symbolic-based multi-agent systems, and a multi-dimension framework.

Agents fail as systems

select a node · run a scenarioswipe · tap a node · run

A multi-agent system with neural and symbolic parts. A task goes to an orchestrator agent, which hands work to a planner, a retriever and an external tool call. The retriever checks a knowledge graph and the tool output passes a rules engine; both are symbolic. A verifier agent combines the results into an answer and can send the orchestrator back to retry. Robustness is measured by perturbing the task, consistency by rerunning the answer, calibration at the verifier’s confidence, and recoverability by injecting a fault at the tool call and checking the retry loop. robustness ×3 calibration consistency ×32 recoverability · retry Task input Orchestrator LLM agent Planner LLM agent Retriever LLM + search Tool call external API Knowledge graph symbolic Rules engine symbolic Verifier critic agent Answer output
ready — pick a dimension and press Run scenario

03 Services

How I can help.

A few ways to work together. Each delivers something usable now, with what you’ll need next built in.

S/01 · Prototypes

Agentic prototypes

For teams that need proof before they commit budget.

Today
A working agentic workflow on your own data within weeks, with reliability measured from day one.
Tomorrow
A model-agnostic architecture that survives model swaps, price changes and new requirements.
S/02 · Strategy

AI adoption strategy

For leadership teams in SMEs and public institutions deciding where AI should go first.

Today
A prioritised use-case map, a feasibility check against your data and processes, and a pilot plan.
Tomorrow
AI Act risk triage for each use case, clear ownership, and a skills plan for your people.
S/03 · Evaluation

Reliability & evaluation

For product and engineering teams running LLM or agent-based systems.

Today
A reliability audit across consistency, calibration, robustness and recoverability, plus an evaluation harness your team keeps.
Tomorrow
Monitoring thresholds and a documented evidence trail your compliance team can map to the AI Act’s accuracy and robustness requirements.
For EU bodies, programmes and consortia, I’m also available as an expert evaluator, advisory-board member and speaker. Get in touch
Regulation-ready, not legal advice. AI Act readiness is built in as technical evidence and regulation-aware design; legal conformity decisions stay with your counsel.

04 Notes

Methodological notes.

Notes from research on what’s changing in AI and what it means for the systems we build.

C:\TRILAR\NOTES

    VIEW

    Same models, different outcomes

    · Strategy

    Everyone has access to the same frontier models. What sets organisations apart is how quickly — and how safely — they turn them into operations people can rely on.

    The AI Act, read as a test suite

    · Regulation

    The AI Act asks high-risk systems to perform consistently throughout their lifecycle. That is an engineering requirement before it is a legal one — a regression test before a legal memo.

    Where symbols beat scale

    · Neuro-symbolic

    The next reliability gains won’t come from scale alone. They’ll come from pairing language models with rules, knowledge graphs and verifiers that make behaviour checkable.

    The demo worked. Then Tuesday happened.

    · Agents

    Agents fail as systems. Small errors compound across steps and hand-offs, so reliability has to be engineered for the whole chain — including how it recovers when something breaks.

    Same prompt, different answer

    · LLMs

    A benchmark score won’t tell you whether a model gives the same answer twice, or whether it knows when it doesn’t know. In production, consistency and calibration matter more — and both can be measured.

    C:\TRILAR\NOTES>

    05 About

    A system of communities, processes and models.

    Portrait of Dr. Jure Trilar
    jure_trilar.webp

    I’m Dr. Jure Trilar, a researcher at the University of Ljubljana’s Faculty of Electrical Engineering where I earned my doctorate. For more than two decades I’ve worked at the intersection of emerging technologies and human-centred design.

    That mix shapes how I see AI: a system of people, processes and models, where adoption fails for human reasons as often as for technical ones.

    My doctoral research developed concepts and tools for designing family-friendly technology, including ways to evaluate interfaces and measure how people perform with digital services. Today I apply the same discipline to AI: measuring how language models and agentic systems behave, and making them dependable.

    Alongside research, I’ve worked inside European AI governance and on the ground with SMEs, public bodies and international consortia in numerous EU funded projects.

    Deep domain mastery

    Policy & standards

    • EU AI regulation, ethics and governance initiatives
    • Board member, national standardisation committee on AI (SIST/TC UMI)
    • Board member, SIST/TC BLC (blockchain)

    Innovation ecosystems

    • Former head, now research & strategic development advisor, EDIH 4PDIH
    • Member, expert council for Slovenia’s EDIH support ecosystem
    • Expert member, RDI Hub advisory group, SPIRIT Slovenia

    Project excellence

    • Lead coordinator, SmartCommUnity (Interreg Alpine Space) - REGIOSTARS 2026 award finalist
    • Project lifecycle management, proposal evaluation and grant writing - Horizon, Interreg
    • Cross-sector coordination - national authorities, public institutions, industry and the public

    Full-stack human developer

    • Web 4.0 - automation and agentic AI development
    • Web 3.0 - Solidity and blockchain, lecturing since 2017
    • Web 2.0 - Progressive web and app architecture
    • Product - UI/UX design strategy and visual identity

    06 Engagement

    Start with a short conversation.

    Tell me what you’re building or deciding. In a 30-minute scoping call we’ll see whether there’s a fit and what a useful first step looks like.

    • AvailableSelect strategic advisory roles, expert evaluation, and keynotes.
    • OpenR&D consortium partnerships & executive AI strategy sessions.
    [Initiate Collaboration]

    How it starts

    1. 01Send a few linesTell me what you’re building or deciding.
    2. 02ScopingWe see whether there’s a fit.
    3. 03A useful first stepYou leave with a concrete first step, whether or not we work together.

    Translate this page

    Machine translation by Google Translate. It loads only when you pick a language.