Backboard.io and MLH Team Up to Advance Stateful AI Development and Shape the Next Generation of AI Builders

• Backboard.io’s AI memory infrastructure expanded support from 2,000 to 17,000+ models within only a couple of months
• Its new partnership with MLH promotes stateful AI development across hackathons and challenges
• CEE developers influence Backboard.io’s roadmap through cost-conscious, efficiency-focused engineering practices
• After memory, Backboard.io targets cost-efficient agentic execution as AI infrastructure’s next challenge

This April, Backboard.io—the Canadian startup infrastructure platform of persistent and shared memory for AI systems across sessions and models—entered into a year-long partnership with Major League Hacking (MLH), the world’s largest developer community. Timed with Global Hack Week, the collaboration aims to equip developers with persistent AI memory capabilities and promote stateful application development as a foundational industry standard. 

Meet the Partners

Founded in 2025, Backboard.io is an AI infrastructure company that provides persistent, portable memory for AI applications and agents. Its platform enables developers to maintain context, preferences, and long-term knowledge across sessions, models, and workflows, overcoming the stateless nature of most LLMs. By combining memory, context management, model routing, and orchestration into a single layer, it helps teams build more reliable and scalable AI systems.

MLH is the world’s largest developer community for tech students and aspiring software engineers. It organizes hackathons, educational programs, Global Hack Weeks, and developer challenges that help participants build technical skills, portfolios, and professional networks. MLH connects students with industry partners, tools, and mentorship opportunities while fostering hands-on learning through real-world projects.

Rapid Infrastructure Expansion and a Partnership to Put Persistent AI Memory in Students’ Hands

ITKeyMedia approached Backboard.io’s co-founder Jonathan Murray to find out more about his company’s progress that led to the team-up with the MLH, its input in the joint initiatives, and its ambitions and expectations:

Last time we talked was only this February and Backboard.io’s model support appears to have grown from ~2,000 to 17,000+ in just a couple of months. What fundamentally enabled this kind of scaling?

Jonathan Murray: Unified layer with BYO API key is the whole trick. Once you stop trying to be the middleman on inference and just give developers one consistent interface to bring their own keys, the model count stops being a moat — it becomes a checkbox. The real work has been everything that sits above the call. Since February, we:

• shipped Adaptive Context Management across all 17,000+ models,
• made state management free, made context management free,
• locked in the MLH partnership,
• and announced our multimodal launch in April/May where image, voice, video, and music are stateful from day one.

At that scale, what are the challenges to maintaining consistent memory behavior across vastly different model architectures and quality tiers?

JM: Our memory layer is model-agnostic by design. It sits above the inference call, so swapping a frontier model for a small open-source one doesn’t change the memory behavior. The harder problem is everything around memory:

• context windows that vary wildly between providers,
• reasoning formats that no two vendors implement the same way,
• billing fields that don’t line up.

We solved the context piece with Adaptive Context Management, which automatically reshapes state to fit whatever model you route to, with a 20% raw budget for live inputs and intelligent summarization for the rest. If summarization can’t fit on the target model, we fall back to the larger source model to compress more efficiently.

Importantly, none of this is possible with a knowledge graph. Knowledge graphs don’t scale — they delete data and call it ‘curation.’ People get wowed by 3D mind-map visualizations, the Obsidian-style stuff, but if you actually understand what you’re looking at, it’s not impressive. Looking cool doesn’t translate into scalable long-term infrastructure.

You planned to launch an open marketplace of tools built on Backboard.io, where developers and research teams can publish assistants, plugins, and workflows that others can fork, extend, and deploy. How’s your progress in this direction?

JM: The marketplace was a good idea, but when we kept asking ‘what can we do to serve our developers today,’ the answer wasn’t a marketplace. It was cost: coding costs, Claude costs, Cursor costs, — that’s the pain about which we hear over and over. So we shifted that energy into a much higher-leverage product, which is launching soon…

We concluded that CEE developers optimize for cost and efficiency over brand loyalty. How has that influenced Backboard.io’s product roadmap or pricing philosophy?

JM: CEE devs aren’t delusional token-maxxers, they build strong products with cost in mind. Compare that to a lot of American devs who are just smashing budget spend and hoping the runway covers it. That discipline is exactly the kind of customer for whom our roadmap is being built, and that’s why we made state management and context management free. If cost discipline is your culture, we want to be the obvious infrastructure choice.

Do you see CEE developers contributing back to Backboard.io as infrastructure co-creators of plugins, tools, frameworks, rather than just users?

JM: Yes, and the framing matters. They’re not contributing because they’re hobbyists looking for clout. They’re contributing because they’re building real products and they want the infrastructure underneath them to be solid. That’s a much healthier dynamic than ‘community engagement’ theater.

Through the MLH partnership, Backboard.io is effectively shaping how the next generation builds AI. Are there any specific bad architectural habits you deem worth eliminating this way?

JM: Single-LLM dependency and memory as a sideloaded feature instead of a first primitive portable foundation. Full stop. We’ve watched major providers change their minds overnight, lock entire teams out of their accounts, impose rate caps with no warning. If your whole product breaks because one vendor changed a policy on a Tuesday, then you don’t have a product, you have a dependency. The other habit is hard-coding context window assumptions and reasoning formats. Every provider does ‘thinking’ differently: OpenAI uses effort levels, Anthropic uses token budgets, Google does both depending on the model, and the outputs come back in completely different shapes. If you build to one provider’s quirks, you’re going to rebuild every few weeks. We want the next generation of devs building model-agnostic from day one, because the providers have shown us repeatedly that they will not protect you.

With 17,000+ models available, might choice become a liability for developers, particularly for younger ones at MLH’s hackathons? How do you guide them toward optimal configurations?

JM: Choice is a liability without measurement. Telling a hackathon dev ‘we have 17,000+ models’ is useless if they can’t figure out which three matter for their specific use case. That’s why we built StackEval. It’s in public beta now at stackeval.io. You drop in your Backboard API key, pick the models you’re considering, choose your task type—Q&A from memory, summarization, RAG, or LoCoMo for long-term conversational memory—and run a live or dry-run evaluation against your actual data. You get cost, latency, ROUGE-L, F1, and BLEU side by side: fastest, sharpest, cheapest, — ranked. Hackathon devs don’t have time to manually benchmark dozens of models, and they shouldn’t be guessing. StackEval turns the 17,000-model menu into an empirical decision in minutes.

Since MLH focuses heavily on students, is there a way to ensure that projects built on Backboard actually graduate into production systems, rather than staying hackathon demos?

JM: The honest answer is the infrastructure has to make graduation the path of least resistance. If a student builds on Backboard.io during a hackathon and then has to rip everything out to go to production, we’ve failed. The whole point of the lifetime free state management offer for MLH hackers, plus free context management, is that there’s no rebuild moment. What they ship at the hackathon is what scales. That’s how you turn demos into companies.

What is the next bottleneck after memory that you believe will define the next phase of AI infrastructure?

JM: Cost of agentic coding is the bottleneck right now and it’s about to get worse as more devs move into agent-driven workflows. The math doesn’t work at current token consumption rates. That’s why our next launch attacks it directly. It’s a CLI we’ve built and have now switched internally for the last three weeks, exclusively. We cut Cursor and Claude Code entirely. Our monthly spend is dropping from USD 44K to roughly USD 5K. The CLI performs at parity or slightly better than Claude Code, while reducing token use by up to 90%. That’s only possible because of our memory architecture and Adaptive Context Management, which no other stack could build. Memory was the first primitive. Adaptive context management was the second. Cost-efficient agentic execution is the third, and they’re more connected than people realize, you can’t solve the third without having solved the first two.

As part of its partnership with Backboard.io, MLH adds a Best Use of Backboard.io track to its 50 hackathons, three Global Hack Weeks, and a dedicated developer challenge, encouraging participants to build AI applications with persistent memory and long-term context. Developers in MLH’s global community will gain free access to Backboard.io credits and lifetime state management for student projects, allowing them to make existing AI tools stateful regardless of the underlying model. Participants who incorporate Backboard.io into their projects will also have opportunities to earn exclusive rewards.

‘MLH is where the most ambitious developers, hackers, and software creators go to learn fast and build the future of technology. By integrating Backboard.io into our global hackathon league, we are giving student developers the foundation they need to build more sophisticated, context-aware AI. Backboard.io’s ability to handle persistent state and model routing through a single API allows our community to focus on what they do best: innovating and shipping real-world solutions,’ MLH’s co-founder and CEO Mike Swift states.

The partnership between Backboard.io and MLH reflects a broader shift in AI development, from building isolated prototypes to creating stateful, production-ready systems capable of retaining context and knowledge over time. By bringing persistent memory infrastructure directly into one of the world’s largest developer communities, the collaboration gives a new generation of builders access to tools that can help transform hackathon projects into scalable products. For CEE developers, whose strengths often lie in pragmatic engineering, cost efficiency, and open-source innovation, the initiative provides an opportunity to compete on a global stage while shaping the future architecture of AI applications.