This recap covers the Open Community for AI track at Open Community Experience 2026 (OCX26), with sessions focused on AI agents, edge AI, domain-specific systems, security, and the role of open source in building production-ready AI systems.
Across three days at OCX26, the Open Community for AI stayed grounded in one question: what actually holds up when AI moves from demo to production?
The focus of these sessions was on constraints, trade-offs, and system design. From small language models at the edge to AI-driven security workflows and data infrastructure, the discussions kept returning to the same point: AI systems fail or succeed based on how they are built, not how impressive the model looks.
Key topics from the AI track
A few patterns kept coming up across the sessions:
- Smaller, specialised models are replacing the assumption that bigger is always better
- AI systems require structure, context, and constraints to behave reliably
- Security, evaluation, and observability are now part of the core system, not add-ons
- Open source is enabling control, interoperability, and transparency across the AI stack
Alongside the main sessions, the track also included Birds of a Feather discussions and hands-on workshops that extended the conversation beyond formal presentations. The BoF sessions focused on practical concerns around AI systems, including topics like agent security and the role of the Eclipse IDE and platform in an AI-first landscape.
In parallel, workshops explored open collaboration across AI ecosystems, with sessions like Collaborative AI: The Open Source Path bringing together initiatives such as MOSAICO, HIVEMIND, and ASTIR. These sessions focused on how projects align architectures, share approaches, and build interoperable systems across organisations. Together, these formats created space for deeper technical exchanges, real-world concerns, and cross-project alignment that do not typically surface in standard talks.
Day 1 at Open Community for AI: Highlights
Small LLMs at the Edge are the engine for open source scalable AI agents with Luca Bianchi
Luca Bianchi focused on a shift that is already happening in practice. “We are coming out from the scaling race”, he said, moving away from the assumption that larger models are always the answer.
The session showed how smaller, specialised models can deliver strong results when combined with the right context and architecture. Edge deployment was a key part of this. Running models locally enables real time processing with ultra low inference time while maintaining privacy protection by keeping data on-device.
The technical focus was on distillation, optimisation, and how to build agents that operate within constrained environments. You can see the recording on our OCX YouTube channel.
One key takeaway: Smaller models combined with the right context and architecture can outperform larger models in real-world systems.
Commit to quality: AI-enhanced testing in open source with Shelley Lambert
Shelley Lambert addressed the impact of AI on software quality, especially in open source ecosystems.
The core issue she highlighted is scale. AI is accelerating development, but it is also increasing the number of issues teams need to handle. As mentioned in the session, “AI is generating 4x more code but is in fact generating 10x more vulnerabilities.”
The session focused on testing, validation, and triage workflows. AI can assist with identifying issues and proposing fixes, but those systems need to be integrated into existing pipelines. Projects like Eclipse Adoptium and Eclipse AQAvit were used as examples of how this can be done in practice. You can see the recording on our OCX YouTube channel.
One key takeaway: AI increases output faster than it improves reliability, so testing and validation must scale with it.
Understanding machine decisions with Haishi Bai
Haishi Bai focused on explainability and the limits of current AI systems. One line captured the problem directly: “We don’t understand the rules that AI creates.”
The session introduced approaches for analysing model behaviour, including systems that move beyond binary logic and support graded reasoning. The goal is not to simplify models, but to make their outputs interpretable enough to support real decisions. You can see the recording on our OCX YouTube channel.
One key takeaway: Without explainability, AI systems cannot be trusted in environments where decisions need to be justified.
Day 2 at Open Community for AI: Highlights
Sharper AI with DSLs: Spec-driven development with Miro Spönemann and Andy Gordon
Miro Spönemann and Andy Gordon focused on reducing ambiguity in AI systems through domain-specific languages. The key shift is moving from prompts to structured specifications. As they mentioned in the session, “specification becomes the source of truth.”
DSLs provide constraints that guide model behaviour. They reduce token usage, improve consistency, and make outputs easier to validate. They also enable a shared layer between developers, tools, and AI agents.
The session connected directly to the Eclipse ecosystem, especially language engineering and tooling. You can see the recording on our OCX YouTube channel.
One key takeaway: Structured specifications reduce ambiguity and make AI systems more reliable than prompt-based approaches alone.
Scale application security with AI-augmented vulnerability remediation with Nitish Tyagi
Nitish Tyagi focused on how AI is reshaping application security. He said, “AI is generating 4x more code but is in fact generating 10x more vulnerabilities.”
The session outlined a system-level approach:
- Use Eclipse SCA Tools and SAST tools to identify vulnerabilities
- Provide structured context to AI systems
- Use AI agents to propose fixes
- Apply human oversight selectively
The goal is not to remove humans, but to use them where they are needed most. You can see the recording on our OCX YouTube channel.
One key takeaway: AI must be used to handle the scale of vulnerabilities it creates, with targeted human oversight.
10 practical learnings from AI projects with Philip Langer and Stefan Dirix
Philip Langer and Stefan Dirix focused on what happens when AI systems move beyond prototypes. The session pushed back on a common pattern: “It’s very easy to create an impressive demo… but you will get into a lot of problems if you just promise to deliver that like this.”
The learnings covered design, development, and production:
- Clear goals matter more than adding AI features
- Context management is critical
- Prompt-based logic breaks at scale
- Evaluation is required at every step
- Monitoring and observability cannot be skipped
The discussion stayed close to real constraints rather than ideal setups. You can see the recording on our OCX YouTube channel.
One key takeaway: AI systems fail in production because of engineering gaps, not model limitations.
Day 3 at Open Community for AI: Highlights
Building AI assistants for DSLs: Experiences and findings from Langium AI with Benjamin Wilson
Benjamin Wilson focused on building assistants that operate within domain-specific languages. The session showed how assistants can be integrated into DSL workflows using a simple process:
- Define the goal
- Select the model and tooling
- Build evaluation frameworks
- Iterate based on results
One point was emphasised repeatedly: “Evaluations are really, really critical… we really should have done this first.” The system improves through iteration, with AI exposing gaps in both the model and the language. You can see the recording on our OCX YouTube channel.
One key takeaway: Evaluation is the foundation for building reliable AI assistants in domain-specific systems.
Eclipse Aidge: opens up the power of edge AI with Axel Farrugia and Iryna de Albuquerque Silva
Axel Farrugia and Iryna de Albuquerque Silva addressed fragmentation in edge AI tooling.
Current ecosystems lead to “fragmented solutions” and “innovation bottleneck[s]” due to complex pipelines and limited interoperability.
Eclipse Aidge brings these steps together into a single framework, covering optimisation, deployment, and hardware integration. The system is built around open source. As highlighted in the session, “the entire code base is open source and is hosted by the Eclipse Foundation.”
This allows teams to adapt and extend the platform instead of relying on fixed solutions. You can see the recording on our OCX YouTube channel.
One key takeaway: Integrated, open frameworks are required to make edge AI deployable at scale.
X-Road 8 “Spaceship”: From launch pad to data spaces with the help of the EDC with Petteri Kivimäki
Petteri Kivimäki focused on data infrastructure and interoperability. X-Road is an open-source solution for secure data exchange, designed to support decentralised systems without central control.
The system is already widely used, with more than 29 countries using X-Road worldwide and almost 600 million potential end users. This moves the conversation beyond individual AI systems to the infrastructure they depend on. You can see the recording on our OCX YouTube channel.
One key takeaway: AI systems depend on interoperable data infrastructure to operate across organisations and ecosystems.
What this means in practice
Across the three days, a consistent pattern emerged. AI systems do not fail because models are weak. They fail when the surrounding system is incomplete.
Structure, context, evaluation, security, and infrastructure are now part of the same problem. Teams that treat them separately run into issues as soon as systems scale. The sessions showed a clear direction. Move away from isolated models and toward integrated systems that can be built, tested, deployed, and maintained over time.
Many of the more detailed questions around system design, security, and interoperability were discussed in smaller BoF and workshop sessions, where participants could challenge assumptions and compare real implementation approaches.
If you missed one of the sessions from the Open Community for AI, you can now see them on our YouTube channel.
