How Zipline Uses Hedgehog for AI Training
Zipline is a drone delivery company that trains AI on private cloud infrastructure to autonomously fly drones and drop packages in precise delivery locations. Florian Berchtold, Zipline’s Principal Engineer responsible for AI developer productivity, highlighted Hedgehog’s crucial role in their operations. Zipline chose an on-premises strategy for their AI infrastructure due to significant cost efficiencies and enhanced governance compared to public cloud options. Florian, a software engineer rather than a network engineer, sought a high-bandwidth networking solution that didn’t demand extensive network CLI expertise. Hedgehog provided a Kubernetes-native, declarative API, allowing Zipline to describe their infrastructure’s desired state in a familiar language, abstracting away complex networking configurations like port channels.
Previously, with a smaller server footprint, Zipline utilized Hedgehog for collapsed core designs, achieving high availability and high bandwidth on a modest scale without requiring specialized networking knowledge. Now, with over sixty servers across multiple racks, Hedgehog continues to be their preferred solution, supporting the larger spine-leaf topology required for their expanded infrastructure. However, a gap existed: while Hedgehog solved the internal fabric networking, Zipline still needed to connect their private cloud to the public internet, necessitating a firewall/router solution. This interim solution involved expensive, commodity legacy firewalls that provided far more capability than Zipline needed for the limited bandwidth they utilized, leading to significant unnecessary costs.
Florian anticipates that Hedgehog’s new Transit Gateway demonstration will fill this crucial gap. He expects the gateway to provide essential routing capabilities, allowing their internal private fabric IPs to access the public internet, along with Network Address Translation (NAT) and basic port forwarding to expose on-premise hosted services. This new functionality from Hedgehog aims to replace their costly existing firewalls, offering a more integrated and cost-effective solution that aligns with their cloud-native infrastructure and declarative management approach.
Presented by Marc Austin, Founder and CEO, Hedgehog, and Florian Berchtold, Principal Engineer, Zipline. Recorded live at Networking Field Day 38 in Silicon Valley on July 9, 2025. Watch the entire presentation at https://techfieldday.com/appearance/hedgehog-presents-at-networking-field-day-38/ or visit https://techfieldday.com/event/nfd38/ or https://hedgehog.cloud/ztlm for more information.
Transcript
Hi, I'm Mark Austin. I'm, um, I'm co-founder and CEO of Hedgehog. And, uh, today, um, we're gonna do, um, three things.
The first thing I'm gonna do just for like 10 minutes is, uh, I'm gonna just set up the problem statement. I'm gonna do that with my good friend Florian here, who's one of our customers. Uh, and then after that, um, Manishh, who's our CTO and Sergey, who's, uh, my co-founder and chief, um, architect are gonna come out and they're gonna do a little bit of review from what we did last summer, uh, around virtual private clouds so that everyone's kind of up to speed.
And then, um, after that they're gonna, um, they're gonna give a demonstration of transit gateway, which was this thing that we promised last summer. Cool. Alright, so, um, just starting off on Zipline.
So Zipline is a logistics company, um, and they deliver things from store to your door. They do it with drones. Um, and they've been, um, piloting for several months in, um, Arkansas with Walmart.
Um, they're live now in Dallas-Fort Worth. I think next market is Seattle Tacoma, where I live. Um, and pretty soon you'll be able to get, you know, deliveries on your door doorstep as well with a drone.
So on that I'd like to introduce Florian Bertold, who is principal engineer for developer productivity at Zipline. Thank you. Mark.
Yeah. So, uh, these, these drones, I think you call 'em zips, right? Correct.
Yeah. So how do they, how do they navigate from store to door? Is there AI involved?
Um, it is, yeah. So after launch, uh, they get packaged, they start off, uh, they completely autonomously fly to their delivery zone, and they use, uh, machine learning models to both identify obstacles during the flight and also to identify perfect location for your burrito to be delivered. Cool.
So your training, your training models? Yes. Correct.
Yeah. Great. And what's your AI infrastructure strategy for?
So we decided to go on-prem for both governance and, uh, cost efficiency reasons. Uh, owning hardware and running them yourself is by far a lot cheaper than, uh, renting in the cloud. And self hosting is not that tricky anymore.
There are good solutions out there. Okay, great. And, um, we talked, I talked about you last summer when I was doing this.
Um, you were at a different company then, right? So, so way back then, what problem were you trying to solve? So back then, uh, we had a a dozen of servers.
So we were looking for high bandwidth, um, networking on a small scale. So we chose, um, collect core type of technology or, or, um, design. And, um, yeah, I wanted to have high availability.
I wanna have at least two switches and was quite tricky to find something off the shelf that, uh, didn't require me, um, getting a license in network engineering. Yeah. And you're a principal engineer for developer productivity?
Correct. Are you a network engineer? No, I'm not.
I'm, I'm a software engineer, so I have a different, uh, tool belt. Um, I don't, uh, I'm not that good at network CLI. Okay.
And so why hedgehog? Like, what, what, what led you to, to us? So, um, hedgehog gave me two things.
Uh, first of all, I was able to describe the, the state of the infrastructure that I want to have in a language that I know which is, uh, Kubernetes. Yep. So I have Kubernetes resources that are declarative, and then I can tell it exactly how it looks like, and all the, the, for me, cryptic logic and, and CLI comments that they need to set up, like, I don't know what a port channel means.
Right. Um, is all handled under the hood. Okay.
So you're at another company now. Yep. Um, why choose why work with Hedgehog again?
I mean, first of all, it worked well and, uh, this time we have about 60, 60 servers that we're working with. Um, so I have multiple racks I need to fill, and the collapsed cord doesn't do it. And I knew that the spine leaf topology, um, is something that you, us support.
Okay. And we, we, you sort of described the problem that we solved for you, but there was always a, a solution gap, right? Yeah.
So what was, what was the gap in the solution that we were providing to You? So, so the whole, the, the main fabric problem was solved. I had hundreds of gigabits between servers.
Um, that's not a problem, but these servers still need to reach out to the public internet, right? So I need some sort of firewall router in between, and that has been missing. Okay.
For now. And so how have you filled that gap in the meantime? Um, commodity, uh, legacy approaches, uh, I wanna call it, um, where we pay enormous fees and licensing and support for getting, uh, two single gigabit high available firewalls that I might just, I need the bandwidth, but I maybe scratch 5% of their capabilities.
Yeah. So why not just keep doing that? Um, it's really a, mostly a cost, uh, problem by paying $50,000 a year for, uh, a gigabit of, of firewall.
Not really something I'm, I'm interested in. Okay. Um, so last question.
What are you expecting, um, Manish and Sergei to demonstrate today? So, uh, from the gateway, I expect to see routing, of course, giving me my internal private fabric, uh, IP subnets access to the public internet. Um, netting of course is involved and, uh, maybe some basic port forwarding so it can, uh, on-prem hosted services make them available, uh, in, in the public internet.
Okay. Great. So on that.
