What happens when cryptography & AI builders spend 10 weeks turning research into products under real constraints? Shape Rotator Demo Day featured private agent infrastructure, self-improving workflows, sovereign data, creator tools, new interfaces + more. See what they built:

Meet @kelrouter, co-founder of @BitRouterAI, one of the teams in the Shape Rotator Accelerator. BitRouter is a self-improving LLM router that works with any harness, model, or loop and optimizes agentic workflows with every run. Production agentic workflows can involve thousands of model calls per run. But not every call requires the most expensive frontier model. Routine tasks such as reading files, formatting outputs, classification, and simple edits can often be handled by lower-cost open models. BitRouter analyzes each call and routes it to the cheapest model capable of completing the task. More complex requests are escalated to frontier models such as Claude or GPT. Rather than making a single static routing decision, BitRouter continuously improves its policy through a reinforcement learning loop: + Act: Route each call according to its intent, complexity, cost, and required quality. + Observe: Track the model, latency, cost, and outcome of every call within an agent run. + Evaluate: Determine whether the selected model cleared the workflow’s quality threshold. + Learn: Incorporate the result into the routing policy so future runs become more efficient. Developers can set the objective they care about, such as cost, accuracy, latency, or a balance between them. BitRouter then adjusts the model mix, automatically escalating difficult tasks or falling back when a provider fails. On Terminal-Bench 2.1, BitRouter reports reducing the cost of GPT-5.5 calls by 32.8%, with a 1.14 percentage point change in solve rate. The team is now expanding its benchmarks across additional SOTA models and agentic workflows. To understand where BitRouter fits in the agent stack, it is also useful to understand what it does not replace: + It’s not just a static gateway: It doesn't just forward calls to a fixed route; it runs an act → observe → evaluate → learn loop that keeps tightening the route as it runs. + It’s not an orchestration framework: It doesn't define your agent's control flow, steps, or state; it routes the calls, tools, and sub-agents your loop already makes. + It’s not an agent harness: It runs under Claude Code, Codex, and the rest, not instead of them. Kelsen studied media and communications at NYU but began teaching himself programming in high school. He met his co-founder Archer Yang through Shanghai’s startup community, and the two began building AI products together before focusing on model gateways and agent infrastructure. Their international background also informs BitRouter’s strategy. Many competitive open models are being developed by Chinese AI companies and offered at a fraction of frontier-model prices. BitRouter aims to make those models more accessible to US developers, not by asking teams to manually select between hundreds of options, but by automatically deciding which model should handle each task. The team’s larger vision is an open, decentralized routing layer for AI with no model lock-in, no agent-harness lock-in, and no requirement to depend on a closed hosted gateway. As agents move from prototypes into production, the question is no longer simply which model is best: it is which model is best for this specific call, and whether your infrastructure can learn to make that decision more efficiently over time. BitRouter is fully open-sourced at github.com/bitrouter/bitrout….