A new startup from ex-Google and ex-Apple researchers is betting that the next AI battleground is not model size, but the feedback systems that make models improve in the real world.
Trajectory has emerged from stealth with a pitch that lands at exactly the right point in the market. According to WIRED, the company was founded by former researchers from Google and Apple and is focused on what it calls the missing layer in modern AI development, reliable feedback infrastructure for evaluation and reinforcement learning.
That matters because the industry has already moved past the phase where benchmark scores alone can carry a product story. Enterprise buyers want to know whether a model behaves well in production, under messy inputs, changing workflows, and repeated user correction, and that requires tooling that can measure, trace, and improve performance over time. In other words, the market is shifting from asking whether AI can answer a prompt to asking whether it can keep getting better after the prompt leaves the demo stage.
The case for this kind of infrastructure is easier to see when you look at how AI systems are actually built today. Models are trained, evaluated, deployed, monitored, and then retrained or tuned again, but the link between production behavior and future improvement is still weak in many teams. That gap is exactly where evaluation platforms, observability tools, and reinforcement learning workflows have started to converge.
WIRED's report puts Trajectory in that intersection, and the timing is notable. AI development has become more industrial, which means teams need tools that can tell them not only whether a model is "good," but where it fails, how often it fails, and what kind of feedback should flow back into training. That is a practical problem, not a theoretical one, and it is becoming more valuable as companies move from pilots to live deployments.
It also helps explain why investors keep circling this part of the stack. Building foundation models is capital-intensive and crowded, while evaluation and feedback tooling can sit closer to the customer's daily workflow. That gives startups in the category a chance to become embedded in enterprise infrastructure, where switching costs are higher and the product becomes part of the operating system for AI teams.
Ex-Big Tech talent is making a bet
The founders' backgrounds at Google and Apple give the company an immediate credibility advantage, especially with buyers that care about technical depth and execution. Former researchers from the major AI labs often bring not just research credentials, but a lived understanding of how large systems fail in practice, which is exactly the perspective this category needs.
This is part of a broader talent pattern that has become hard to miss. As the AI market matures, more ex-Big Tech researchers are leaving the labs to build the plumbing around AI rather than the models themselves. Reuters and other outlets have repeatedly shown how the most aggressively funded corners of the market are moving toward infrastructure, agent tooling, and systems that make AI more reliable rather than simply more impressive.
That shift is visible in the broader observability market as well. Companies such as Arize, Galileo, and Braintrust have already helped define the space around monitoring, testing, and evaluating model behavior, and Trajectory now appears to be entering a category that is still early but increasingly crowded. The real opportunity is not just tracking failures, but creating a loop that turns those failures into measurable product improvement.
For SF readers, the bigger story is what this says about capital allocation. If the first wave of AI investment was about building the model, the next one is about proving the model works in the real world, then tightening the loop so it gets better with use. That is a quieter bet, but it may be the more durable one. The companies that can turn feedback into advantage are likely to matter more than the ones that simply chase the biggest parameter count.
Trajectory's emergence also reflects a simple truth about this market. The more AI is used in customer support, coding, search, and enterprise workflows, the more valuable it becomes to know exactly how it behaves under pressure. That is where evaluation stops being a nice-to-have and starts looking like core infrastructure, which is why this startup's pitch feels less like a niche product launch and more like a signal about where the next layer of AI value may form.
Also read: Cognition's funding talks show how far AI coding bets have run • China is locking down its AI talent as the race with the US hardens • Robinhood lets AI agents trade stocks and pay by credit card
