Trust, Openness, and Graphs: Three Patterns from ODSC East 2026 

ODSC East 2026 was a busy week of workshops, talks, and book signings — and a great one. If you get the chance, we strongly recommend catching as many replays as you can, since we won’t walk through every talk we attended here. Instead, we want to focus on three patterns that emerged across the week. 

We have moved past the LLM boom 

Yes, models continue to improve — some faster than others. But the most interesting talks were not about models; they were about everything around the models. We are happy to see trust, evaluation, security, and cost move to center stage. 

Jonathan Bown’s workshop on going from PoC to prod with MLflow 3.0 captured the mood well. In it, he walked attendees through tracking models, experiments, prompts, and LLMs in one place (emphasis on that last part!). Breaking the unnecessary barrier between Machine Learning and LLM-based AI is a huge win in our book: ML teams are used to tracking performance and quality metrics, while the rush to ship AI has given LLM-based efforts little time to adapt and settle on the right metrics to track (and how to track them). As we discussed in our thoughts on OptimizedAI, this kind of tracking stops being a nice-to-have the moment agents start running in production and stakeholders ask who approved what. 

Wenxin Du’s talk on architecting secure data access for AI agents using MCP (model context protocol) and OAuth approached the same problem from a different angle. Once agents can act on your behalf, security issues arise faster and are harder to track and resolve. She proposed an approach that treats the agent as the least trusted identity in the stack, with credentials, connection details, and user identity all kept outside the model’s context. Tools expose narrow, parameterized SQL rather than raw query execution, and identity fields are injected from a verified OAuth token at runtime rather than supplied by the agent itself. Her demo used Google’s MCP Toolbox, but the pattern itself is tool-agnostic and worth borrowing. 

The bottom line: people are no longer asking whether the technology works — they are asking how to make it trustworthy enough to deploy. 

Open source is dead! Long live open source! 

AI really has been a double-edged sword for open source. On the one hand, repos (repositories) are getting bombarded with agent-initiated PRs (pull requests), with some agents even rebelling against the codebase owners when their code is rejected. Many developers have noticed a decline in code quality. Some open source projects have even had to switch to closed source because they could not keep up with code review at “agent-swarm” scale. On the other hand, teams who have learned to use the tools well are getting faster. This is an over-simplification of the issue, but one thing was clear at ODSC: agents have not killed OSS (Open Source Software), contrary to what developers might have been hearing on social media. 

Beyond our earlier mentions of MLflow and MCP itself, there were many more examples of OSS winning. A talk by Preset showed recent developments in Apache Superset, an open source data exploration and visualization platform. We were happy to see OSS analytics platforms still getting love — Preset’s platform, built on top of Superset, unlocks AI-native analytics. At the end of their talk, they also unveiled agor, a shared canvas to track and orchestrate agents that seemed very promising. For us, it looked like the kind of UI that makes the agent story click for people who would otherwise glaze over. 

One last note on open source: though cost and transparency are usually framed as the advantages of OSS, Ivan Lee from Datasaur made the larger case for what he called a sovereign LLM stack. Take a top open-source model, deploy it on hardware you control, fine-tune it on your own data, and you stop renting your intelligence from someone else. Your sensitive data never leaves your servers, your prompts are not training a competitor’s next-generation model, and a vendor sunset cannot pull the rug out from under a fine-tuned system you spent a year building. The blind-test parity between open and closed leaders is what makes this a real option in 2026 rather than a compromise, and he showed that the cost of developing this kind of stack is much closer to using frontier models than one might think — only this time, the stack is yours. 

Graphs are useful… when used well 

We cannot talk about ODSC without talking about how the graph community really showed up! 

Amy Hodler and David Hughes’s talk on mmGraph is the one we will be thinking about for a while. The idea is to fuse hypervectors with your knowledge graph so agents can reason with both symbolic and learned context. That framing alone was worth the price of admission. Grounding an agent in a structured representation of the world rather than a sliding window of tokens lines up with what we have been pushing for in client work for years. Hypervectors give the graph a fuzzy, learned dimension that sits alongside the precise, symbolic one, and we think this hybrid pattern is going to matter more and more over the next couple of years. 

A separate session on the compounding power of connected data made the business case from the other direction. Denise Cosnell, Ph.D., walked through what disconnection costs an organization: lost time, repeated work, decisions made on partial information. Anyone working in our field feels these costs intuitively, but quantifying them is another matter. Putting concrete figures on the table is something we are taking on ourselves, so watch this space. 

Of course, this enthusiasm comes with a qualifier. Graphs are useful… when used well. Throwing a graph database at every problem will not save you (sorry!), and we have seen plenty of projects where a well-designed relational schema would have done the job with less ceremony. We see the inverse just as often: teams forcing “graph-y” problems through a relational schema, then paying for it in painful joins, brittle queries, and infrastructure bills that would shrink if they had picked the right tool from the start. The teams we are seeing pull ahead are the ones modeling their domain carefully, picking the right entities and relationships, and connecting the graph to the rest of their data fabric rather than treating it as a parallel universe. The community at ODSC felt sharp and pragmatic on this point — proof that the conversation around graphs has matured. 

Speaking of community — a special thanks to GraphGeeks and Glasswing Ventures for co-hosting the mixer after day two. The ideas zooming around the room and the warmth of the crowd made it the social highlight of the week. If you missed it, put it on your list for next year! 

Final thoughts: keep your stack interchangeable 

If there is one thing the week reinforced for us, it is that the individual pieces of your stack need to be interchangeable. 

New tools land every day. The model you fine-tuned last quarter is mid-tier this quarter. The vector store you picked will have a faster competitor next month. The agent framework you committed to last week has forked twice since then. We see this pattern often: a team treats one of these as load-bearing, only to end up spending more time on migrations than on the work that actually moves their business forward. Interchangeability is not free, though, and it is certainly not a single product you can buy off a shelf. It is a discipline that touches your data layer, your retrieval, your model serving, your evaluation harness, and your orchestration. Done well, you can swap a piece without rewriting the rest. Done poorly, you end up with abstraction sludge that is worse than the lock-in you were trying to avoid in the first place. 

If you want to talk about what this looks like in practice, that is the kind of conversation we love having. 

Graphable specializes in applied AI, graph databases, and data engineering. We help organizations move from AI experimentation to production-grade systems that deliver real business outcomes.