The technical stage in the Mozart hall, laptops-open workshops and demos that actually run. For engineers, data specialists, CTOs, researchers and security people.
The rest of the festival talks about adopting AI. Here we talk about how it is done: what runs in production, what a thousand requests cost, why an evaluation fails, where an agent system breaks on the third iteration.
We do not accept product pitches dressed up as technical sessions. If a talk has no code, no latency and cost figures and no system running in front of the room, it does not go into this strand. It can go into partner content, clearly labelled as such.
This list is the proposed programme for Edition 01. It will be finalised in December 2026, as talks are confirmed.
What an agent means in production rather than in a demo: orchestration, tools, memory, recovery after failure, the control loop, what happens on the third iteration. Cases where agents are not the answer.
Document chunking, hybrid retrieval, reranking, proprietary data, documents in Romanian, what to do when the knowledge base contradicts itself and how to measure whether an answer is supported by its source.
How to build an evaluation set for your own case, what you are really measuring, model-as-judge and its limits, regressions between versions, how to decide that a change is an improvement.
Latency, throughput, prompt caching, batching, quantisation, choosing between a hosted model and your own. Real cost per request, not estimates.
What it means in practice to keep data and models in the country or in the EU, where processing happens, what European providers offer, what compute costs and when it is worth buying rather than renting.
Prompt injection, exfiltration through tools, the limits of filters, adversarial testing of agent systems, what an incident means in an AI system and how to document it.
From raw data to usable datasets: cleaning, labelling, versioning, usage rights, personal data, Romanian-language datasets and why their absence is a market problem, not a technical one.
When fine-tuning makes sense and when it does not, parameter-efficient fine-tuning, distillation, what you can realistically get from a small open model compared with a large hosted one, licences and what they allow you to do.
Models that see and hear, on-device processing, AI in industry and in the city, robotics. What can and cannot run on cheap hardware, in real conditions.
Want to propose a talk? The strand is editorial: talks are chosen for what you can show, not for the company you come from.
The Smart City Romania network, run by ARSC
Everything you need to know about Romania's fastest-growing industry