How Mandel AI runs a complex, multi-region AI stack on Northflank

Mandel AI helps manufacturers and distributors manage the operational work involved in moving goods. The product ingests information from communication channels, primarily email, and feeds it into AI agents that keep track of promises, requirements, services, and issues.

Behind that relatively simple description is a fairly complicated infrastructure footprint. Mandel self-hosts much of its stack, running its core application alongside workflow orchestration, databases, authentication, observability and communication services, while also managing customer data residency requirements.

Mandel initially ran on Fly.io. As the company grew, infrastructure reliability and regional constraints became increasingly difficult to work around. The team moved to Northflank and now runs its application infrastructure in its own Azure account using Northflank BYOC.

Before Northflank, Mandel primarily ran its infrastructure on Fly.io, where the team started encountering reliability problems across several parts of the infrastructure.

Scaling could be unreliable, machines could disappear unexpectedly, disk performance could be slow, and networking could become unstable.

More difficult were regional outages affecting stateful services. Because volumes were tied to infrastructure in a particular location, temporarily moving a stateful workload elsewhere wasn't necessarily an option. In some cases, the alternative would have been a multi-hour recovery and redeployment into another region.

For Mandel, that translated directly into application downtime and customers unable to log in.

“We had trouble scaling reliably. Machines would die at random, disks would be slow, networking would be flaky. We ended up dealing with issues that were completely outside of our hands.”

Tsvetomir Bonev, Founding Engineer at Mandel AI

Mandel's requirements for a replacement were relatively straightforward: reliable infrastructure, good visibility into what was happening, and enough control to operate the application without constantly fighting the underlying platform.

“We wanted good visibility into what was going on, easy control, and for the things we relied on to work reliably.”

Tsvetomir Bonev, Founding Engineer at Mandel AI

Mandel ultimately deployed Northflank BYOC into Azure.

With Northflank BYOC, Mandel retains its underlying Azure environment while using Northflank as the platform for deploying and operating the workloads running inside it.

That becomes particularly useful given how much infrastructure Mandel operates itself.

Mandel deliberately self-hosts a significant part of its stack.

Customer requirements are one reason. Mandel handles customer information, and introducing additional third-party providers can create questions about where that data is going and who has access to it.

The core Mandel application is written in TypeScript, while much of its LLM processing runs through a self-hosted instance of Windmill, which the team uses for workflow orchestration.

Around that, Mandel operates a collection of other services including:

• MongoDB as its primary database
• Postgres, including a Northflank Postgres add-on used by Windmill
• Redis
• a self-hosted authentication stack
• Langfuse
• EmailEngine
• other services used to connect Mandel with customer communication channels
• custom Northflank add-ons

Mandel needs to understand what is deployed where, how services connect, which infrastructure belongs in which region, and how to reproduce an environment when another deployment is required.

Mandel initially operated from a single region.

As the company gained traction in the US, customers began asking for their data to remain in the United States. Mandel therefore needed to reproduce its existing environment in another region rather than simply add more capacity to its existing deployment.

The team turned its Northflank infrastructure into a template. With help from Claude Code and the Northflank skill, Tsvetomir was then able to use that template to recreate the environment in a new region.

“I made a template with Claude Code using the Northflank skill. It was quite trivial to set up. At the end, I stood up an entire new region by clicking a button, and everything worked.”

Tsvetomir Bonev, Founding Engineer at Mandel AI

Instead of a collection of services that would need to be manually reconstructed, Mandel now had a repeatable definition of its infrastructure.

That makes regional expansion a much more manageable engineering problem. The team can reuse the same infrastructure definition rather than reconstructing a growing collection of services and configuration whenever it needs another deployment.

For Mandel, that is particularly important because expansion into a new geography can also come with customer-specific data residency requirements.

As Mandel has grown, its infrastructure requirements have become increasingly specific.

One of the things Tsvetomir has found most useful about Northflank is how often the platform already supports those requirements.

“Usually I'll think, ‘It would be really useful if I could do this,’ and then find out that I already can. There hasn't been a case where I've really needed something and it didn't exist.”

Tsvetomir Bonev, Founding Engineer at Mandel AI

That has included relatively advanced networking requirements.

For example, Mandel needed to configure DNS overrides globally so that its services could resolve the private endpoint used to connect to MongoDB. The capability was already available within Northflank.

And according to Tsvetomir, this hasn't been an isolated experience.

“We've had this happen probably ten times. I'll have some specific infrastructure requirement and find out Northflank already has a way to do it.”

Tsvetomir Bonev, Founding Engineer at Mandel AI

As Mandel's architecture becomes more complicated, the team can continue using the same platform rather than adding another tool or building its own solution every time a new requirement appears.

It also means Mandel doesn't have to choose between the simplicity of a managed platform and access to more advanced infrastructure capabilities as it grows.

The platform covers most of what Mandel needs day to day, but there are inevitably cases where the team encounters an unusual configuration or infrastructure issue.

In those situations, Mandel has direct access to Northflank engineers.

Rather than working through layers of generic support, the team can get the relevant engineer involved to investigate the problem.

“Once the relevant engineer is involved, I've always received a prompt resolution.”

Tsvetomir Bonev, Founding Engineer at Mandel AI

Mandel expects its infrastructure requirements to grow quickly as customer usage increases.

The company already operates a substantial number of self-hosted services, multiple databases, LLM workflows, private networking and infrastructure across multiple regions.

What has stood out to the team is that increasing complexity hasn't forced it to move away from Northflank or start assembling its own platform.

Instead, as Mandel has needed more advanced infrastructure capabilities, it has repeatedly found them already built into the platform.

“Northflank is incredibly feature complete. There hasn't been a case where I really needed something and it wasn't there.”

Tsvetomir Bonev, Founding Engineer at Mandel AI

Northflank gives Mandel one place to define and operate its services, databases and supporting infrastructure inside its own Azure environment. Templates give the team a repeatable way to reproduce that architecture across regions, while direct access to Northflank engineers gives it a path for resolving the more unusual infrastructure problems that inevitably come with a growing platform.

For a small engineering team running an increasingly sophisticated AI application, that's the main benefit: Mandel can keep adding infrastructure complexity without having to build the platform layer itself.