COMPANY KICKOFF EVENT

Systems of Attention: Inside the Launch of Signal Labs

You can feel it when something genuinely new walks into a room. Not a product. Not a feature. A whole new category.

You can feel it when something genuinely new walks into a room. Not a product. Not a feature. A whole new category.

That's what the Institute felt like on Wednesday night. Operators, scientists, investors, and enterprise leaders had come in for the public launch of Signal Labs, and what got introduced was the company's idea of Systems of Attention, the architectural layer that's been missing from the modern enterprise stack.

Three speakers carried the evening. CEO and Founder Rajeev Ronanki opened with the category itself. Then Jeff Klebanoff, GM of Consumer Markets, took it into rail, logistics, and freight. Plamen Petrov, CTO and GM of Trust Zones, closed by walking through the math underneath the whole thing.

Rajeev Ronanki: A New Layer for the Enterprise

Rajeev started somewhere you don't usually start a launch keynote. 3400 BCE. Cuneiform. Five thousand years later, ENIAC. Then a quick march through the modern enterprise stack: Systems of Record, Engagement, and Insight. The point of the arc was simple. Every layer we've ever built has been about extending memory and decision-making outward.

AI is now woven through all three. And yet, 52% of the Fortune 100 from the year 2000 are no longer around today. Something is still missing.

That something, he argued, is attention. Institutions run on signals from machines, people, markets, operations. Those signals get trapped in silos, drowned out, or just show up too late to matter. 100+ tools per enterprise. 60% of signals arrive too late. 95% never reach action.

To make the case for why this is the right problem to solve right now, Rajeev reached for two unlikely bookends. Herbert Simon, in 1971: "A wealth of information means a dearth of something else, the attention of its recipients." Then Attention Is All You Need, the 2017 paper that gave AI a mathematical way to figure out what matters. Machines got their answer. Institutions are now standing where AI was almost a decade ago.

The fireside that followed pulled the analogy tight. Dr. Adam Gazzaley of UCSF, one of the leading researchers on how human attention actually works, talked through what neuroscience has learned about the brain handling competing inputs. Focus gets allocated by relevance, urgency, and context. Push past those limits and performance falls apart. The takeaway hung in the air. What the brain does biologically, every institution now needs to do architecturally.

Then came the platform. SignalOS™ sits on top as the executive experience layer. Underneath is the SignalGraph™ Blackboard, the coordination intelligence engine that does the actual work. Four primitives hold it together: Signal Ingestion, an Equilibrium Engine, Attention Allocation, and Institutional Memory.

Rajeev got concrete with a $4,200 healthcare claim. In a normal enterprise stack, the warning signs around that claim live in five different systems and never meet. Actuarial drift sits in one place. Pre-auth spikes in another. Upcoding patterns and length-of-stay anomalies somewhere else. Inside SignalGraph they form a single hypothesis cluster. Confidence aggregates. Attention routes days before the claim ever shows up at the door. And here's where it gets uncomfortable. That one $4,200 claim sits inside a $2.3M provider pattern, a $140M plan-level exposure, and an estimated $1.2T of market waste. Roughly 30% of US healthcare spending.

He closed his portion with the launch numbers. $25M in signed contract value, 50+ active client conversations, Lightspeed Venture Partners as lead investor. Attention infrastructure, he said, isn't a feature. It's the next platform layer.

Jeff Klebanoff: Rail, Logistics, and Freight

Jeff was up next, and his job was to show the category travels. So he took the room into a completely different operating environment. Every rail corridor in the country runs on dozens of systems that don't talk to each other. Carriers, terminals, warehouses, shippers. Each one guarding its own data. When something fails across them, no single system sees the cascade until it's already too late to do anything about it.

The numbers reset the stakes.

• $800B+ in U.S. rail and freight at risk every year.
• 100+ disconnected systems per supply chain.
• A 2 to 6 hour early warning window before a cascade becomes irreversible.

Jeff walked through how a real failure plays out. Bearing defect, crew timeout, storm window, chassis shortage, dock bottleneck, SLA breach. Equipment plus weather plus crew plus paperwork, compounding invisibly across 300 miles of corridor. Then he made the point that landed hardest with this audience. The reason nobody connects these signals isn't a technical problem. It's a trust problem. SignalOS gets around it by acting as a neutral coordination layer that no single party owns.

In a representative deployment, SignalOS pulls from nine independent sources. Equipment health, yard capacity, crew scheduling, weather, warehouse management, EDI, GPS, and a few more. The math separates signal from noise. Then ranked intelligence gets routed to whoever actually needs to act on it, hours before any individual system would have caught the problem on its own. Same architecture as the healthcare demo. Different industry. Same outcome.

Plamen Petrov: The Mathematical Backbone

Plamen closed the talk by getting under the hood. His section answered the question some of the most technical people in the room had been sitting on the whole evening. How does any of this actually work?

He unpacked Fisher Information as the way SignalOS measures how much actionable content a signal really carries. He worked through noisy-OR confidence aggregation, the method for combining signals from independent sources without overcounting them. And he covered semantic linkage, which decides whether two signals belong to the same hypothesis cluster or just look similar on the surface.

The architecture itself got its own moment. SignalGraph, at its core, is a blackboard system. It's a coordination pattern that goes back decades in AI research, where independent knowledge sources, detectors, LLMs, rule engines, ML models, human reviewers, self-select into the work based on what they're actually competent to evaluate. Confidence builds. Sufficiency thresholds get crossed. Attention routes.

The detail mattered because it explained something the demos couldn't quite. A System of Attention isn't something you can stitch together from pieces lying around. It needs its own architecture, its own math, its own substrate. Plamen and Rajeev are co-authoring a forthcoming technical paper on it.

What Comes Next

Rajeev kept coming back to one line as people started to file out. The agentic AI era doesn't reduce the need for human attention. It makes attention the scarcest resource in every institution that uses it.

Signals are perishable. Attention is finite. Memory compounds.

Apex, the first generally available release of SignalOS, went live on April 24, 2026. That's when Systems of Attention stop being a category definition and became deployed infrastructure.