David Bensoussan

Separating OS and Application Updates

Tue, 08 Sep 2026 10:00:00 +0200

A process restart takes a few seconds and only loses its own memory. A reboot takes the whole edge server down: the tunnel to the hub, the local Kubernetes node, everything the robot depends on it for. Build one pipeline for both and size it for the reboot, and every small app fix now waits on the same caution a reboot needs, for no reason.

This builds on updating robots in the field and the reference architecture. Read those first if you haven’t.

Platform Adoption: Tooling, Training, Buy-In

Tue, 01 Sep 2026 10:00:00 +0200

Article three covered the architecture: the hub, the edge servers, the GitOps pipeline, the signed OS images. Building that and getting people to use it are two different jobs.

This builds on what a fleet platform costs and the reference architecture. Read those first if you haven’t.

Nothing changes on its own once the platform is live. Engineers keep SSHing into a site and pulling code by hand, even after you tell them the old way is over, because that’s the workflow they know and trust. Getting them onto the new one takes a transition period, documentation, more than one training session, and one-on-one conversations to answer questions. Only after all that does SSH access get turned off.

A Reference Architecture for Robot Fleets

Tue, 25 Aug 2026 10:00:00 +0200

The last article listed five capabilities and the order to build them in. Before building anything, we must agree on the architecture: what runs where, and which direction each connection travels.

The architecture has three layers. The robot runs its own ROS2 stack as before. Above it, one or more edge servers per site carry updates and data between the robot and the hub. Above the edge servers is a single hub, shared by the whole fleet, and every edge server connects to it. The hub never opens a connection to a site, that’s what makes control and security possible.

What a Fleet Platform Costs

Tue, 18 Aug 2026 01:00:00 +0200

The last article said you need a DevOps engineer, in-house or freelance. Before you take that to your CEO, you need to know what it costs.

You must be able to tell him how long he should pay this person before seeing any result. In this article, we will dive into this.

We previously listed five topics a fleet platform gives you: speed, control, visibility, security and compliance.

blog.bensoussan.de

Figure 1. The same three robots at every stage, with one capability added each time. Before, nobody knows what is running on any of them. By the end, the fleet is reachable, reporting, protected and auditable. The pips under each name are the months that stage takes.

What the five steps cost

Step	What you can do once it’s done	Effort	Who carries it
Speed	Rebuild any version on demand, install it with one command	A month	An engineer you already have, part time
Control	Reach every site remotely, change the robot app there by editing Git, ship a versioned OS image, roll either back	3 to 6 months	Full time
Visibility	See a robot degrading before the customer phones	3 months	Full time
Security	Only builds you approved run; revoke one key remotely	6 months	Full time
Compliance	Answer what ran where, and when from your office	3 months	Full time

Step 1 is small enough to start without approving a budget. You just need pipelines turn green. For that, you need to understand how to build the software and train employees to use the CI/CD workflow.

Updating Robots in the Field

Tue, 11 Aug 2026 10:00:00 +0200

Ask a robotics company, especially a young one, how software gets onto a robot that’s already at a customer site, and the answer usually is someone driving there, accessing the robot, pulling the code and getting it to run, watching the robot move for a minute, then moving on to the next one.

The reason is lack of awareness rather than carelessness. How many robotics companies have a DevOps engineer on the team? Most are led by roboticists and researchers who can make anything happen on a few robots. So a great deal of sophisticated software gets written, and then moved around manually.

PROVEN: Six Checks Against AI Slop

Wed, 05 Aug 2026 00:00:00 +0000

feat(transcription): automatic video subtitle generation via faster-whisper #30514 — closed · 17 commits · 3,035 tests passing · 0 lint errors · auto-closed:llm

My AI-assisted pull request to immich, an open-source project, was closed within minutes today because an LLM was involved. Their policy forbids it:

We ask you not to open PRs generated with an LLM. We find that code generated like this tends to need a large amount of back-and-forth, which is a very inefficient use of our time. We cannot have any confidence in an LLM, so if it’s non-trivial for us to verify it works and you don’t show that you fully understand all implications of the change, reviewing the PR is not worth our time.

Zero Overhead

Tue, 04 Aug 2026 00:00:00 +0000

I’ve never kept a blog. Every time I considered it, the overhead argument won: a platform to maintain, a hosting bill, and posts that would have to be written from scratch. Not worth it.

Two things changed this year, so the math finally works.

Who I am

I’m David Bensoussan, a software engineer. Most of what I build outside work lives in one monorepo: the apps, the infrastructure that runs them, and the CI in between. The infrastructure part is a homelab — two second-hand Debian machines running about 45 services: photos, media, documents, monitoring, an LLM server on an old GPU, and the CI runners themselves. The full topology and progress log is at homelab.bensoussan.de if you want the inventory. Everything is provisioned with Ansible, and every significant decision gets written down.