Probably is building a new kind of AI agent, one designed for a world where users need to trust that the AI outputs are reliable enough to drive critical decisions. The company was born from a realization: LLMs can fabricate numbers and misinterpret data, but when organizations rely on those outputs for high-stakes strategy decisions, the consequences compound. To solve this, Probably pairs AI reasoning with a deterministic, local compute engine that forces the LLM through rigorous validations before the user sees the output…

Instead of sending sensitive datasets to the cloud, the company runs its AI agent locally on the user’s machine, keeping the data plane fully contained while using the cloud only for orchestration and coordination. Mathematical work is delegated to a processor-optimized compute engine on the local machine, not the LLM, which is what makes the outputs verifiable.

Probably’s control plane runs on DigitalOcean’s AI-Native Cloud, a fully integrated platform purpose-built for AI workloads from infrastructure through inference. On it, the company has built a privacy-first analytics system that scales without the operational complexity or cost premium of hyperscaler providers. The control plane runs on App Platform within DigitalOcean’s Core Cloud layer today, and as adoption grows, Probably plans to expand into Managed Kubernetes for orchestration and the Inference Engine for open model workloads, all without leaving the platform it already runs on.

The results Probably achieved on DigitalOcean:

• 25% reduction in infrastructure costs: Lower spend than equivalent Amazon Web Services (AWS) configurations, with no metered egress or storage surprises.

• API setup in a day and a half: From zero to production-ready API infrastructure on DigitalOcean App Platform, a timeline Peter Elias says would not have been possible on AWS.

• Tens of thousands of requests per second: Handled through virtual machines running the company’s custom-built online evaluation system through DigitalOcean’s edge network.

A Cloud for Local-First Architecture

Probably’s architecture is unusual. The agent runs locally on each user’s machine while the cloud serves as a coordination layer across the public internet. That creates specific demands on a cloud provider: the platform needed to reliably serve distributed client connections from many different user environments and absorb bursty traffic patterns without pulling a small team away from product development. A hyperscaler would introduce unnecessary complexity and data gravity for a team that deliberately keeps data local. And a neocloud or standalone inference wrapper would only cover part of the stack, leaving Probably to stitch together the rest.

Peter evaluated his options from a practitioner’s perspective. Before founding Probably, he built distributed systems and data infrastructure at Condé Nast and Optimizely.

“I have a lot of opinions about cloud providers and cloud infrastructure,” Peter says. “If it exists, I’ve used it.”

Most platforms fall into one of two camps. They offer raw capability with significant operational overhead, or they offer simplicity at a high price.

“Traditionally, if you were going to build a cloud-heavy product, your options were to spend a ton of time messing around with AWS or Google Cloud Platform (GCP) directly, or using something like Heroku, which was incredibly expensive,” Peter says. “For the stage that we’re at now and for the foreseeable future, DigitalOcean is an attractive option because we get the best of both worlds.”

DigitalOcean’s AI-Native Cloud spans the full picture, from core compute and networking through inference, on a single integrated platform. Wide-area network performance was especially relevant given the architecture. With agents running locally across many user environments, a single session can generate significant load on the coordination layer.

“If you spawn 50 agents at once…that’s going to generate potentially hundreds of requests per minute just from a single client,” Peter says.

As concurrent sessions scale across a growing user base, the cloud layer needs to absorb traffic spikes without degradation. The path from simple to complex infrastructure also played a role in the decision. Most providers force teams to choose between ease of getting started and the ability to handle sophisticated workloads later.

“DigitalOcean’s the only provider that actually gives you that whole scale. There’s no equivalent tier with any other provider where you can start simple and then scale up complexity. That’s a really unique positioning, and it’s worth a lot,” Peter says.

Running the Control Plane on DigitalOcean

Probably runs its core API infrastructure on DigitalOcean App Platform, a fully managed platform that handles deployment, scaling, and infrastructure so a small team can stay focused on product. App Platform is the backbone for agent orchestration and billing, with built-in autoscaling to absorb the bursty traffic that Probably’s architecture generates and dedicated IPs designed to support secure and compliant connections to external services.

Critically, the architecture keeps customer data entirely on the customer’s machine and within their network. Cloud LLMs only power the reasoning and orchestration; the data itself never leaves the user’s device. Requests sent to external LLM APIs contain only the information required for planning and reasoning.

“DigitalOcean hosts two very critical systems for us. One is billing, and the other one is the evaluation system for agent performance,” Peter explains.

The evaluation system is worth particular attention. Probably built it to meet the unique challenges of assessing a data agent’s behaviors at scale, and the system has proven it can handle real production load.

“Our agent evaluation system, built from scratch to meet the challenges of evaluating a data agent, ingests traces through the DigitalOcean edge network. We have seen high watermarks reach tens of thousands of requests per second, handled by modestly sized virtual machines.”

Traffic flows through a CDN layer before reaching DigitalOcean Load Balancers, creating an edge-optimized architecture that keeps costs low while absorbing bursty workloads. The platform also supports authentication and user management, making it the company’s primary system of record for operational data.

25% Lower Costs, Faster Time to Production

Since launching early production pilots in mid-to-late 2025, Probably has seen customers actively using the platform with encouraging outcomes.

“We have about five or six of those active pilots who are now pretty actively using the software, and they’re getting good results,” Peter says.

Behind that progress, DigitalOcean has reduced both cost and operational drag. The savings are concrete and quantifiable.

“When you factor in disk and network metering on AWS, we save at least 25% for the same hardware configuration without having to reserve capacity. With DigitalOcean, we have fixed costs, more value, and less setup time,” Peter says. “For most startups, that’s a better deal than trying to optimize AWS spend when you should be focused on product.”

Speed to production has been just as valuable. Setting up Probably’s initial API on DigitalOcean took a day and a half. Peter noted: “That doesn’t happen on AWS.”

For a small team building a technically demanding product, the quality of customer support also changes what’s possible day to day.

“Anytime I have the slightest problem, I can just talk to the people at DigitalOcean on the engineering team, and I get a great answer. That kind of responsiveness and personal touch—you can’t get that from a hyperscaler. You’ll just get buried in a support queue for the rest of your life, and especially as a startup, that’s the last thing you want,” Peter says.

Scaling Into the Inference Era

Probably continues to refine its system while preparing for broader adoption. With a growing waitlist and increasing confidence in its architecture, the team expects to grow into more of the DigitalOcean AI-Native Cloud, moving onto DigitalOcean Kubernetes for orchestration and DigitalOcean Inference Engine for open model workloads, without leaving the platform they already run on.

“As the company grows, it’s going to be important for us to have inference capacity pools that are diversified. With DigitalOcean, we’ll be able to easily deploy open models onto our existing platform, which is less work for us than juggling a bunch of different providers,” he explains.

The broader opportunity is sizable. Most organizations want to rely on agents to perform critical data analysis and assist in strategic decision-making. But some are learning that the current generation of frontier models are not always reliable for fact-intensive work, and that verifying their outputs can remain challenging. Probably is building toward a world where any individual or corporation can interrogate their most critical data directly and get verified answers they trust enough to act on. That the data stays on the user’s own machine is designed to help mitigate a privacy concern that has kept many enterprises from adopting AI data agents at all. Getting there means scaling inference predictably as adoption grows, which is exactly what the DigitalOcean AI-Native Cloud is built for.

“With DigitalOcean, you’re going to get affordability. You’re going to get scale and capacity and headroom for expansion,” Peter says. “But most importantly, you’re going to save the most time.”