Employment work
Rely
Frontend Lead & Product Designer
2025production
Next.js, Go, WebSockets, Tailwind v4
Factories use Rely to check every part on the line. Operators need a verdict they can read at a glance. Engineers need to set up what gets checked and train the models behind it. I designed and built the frontend for both.
Decisions
1. Rely first: production and training stay in sync
Rely runs on the line. Rely Cloud is where we label images and train the models. Some clients have no Rely Cloud, some run it on their own servers, some use it in the cloud, and some are custom projects with their own interfaces. I had to design one link that worked for all of them.
I decided production leads: you don't think about the cloud until you use the line. A station is one inspection point on the line. When you create one in Rely, its project in Rely Cloud is created too. A station has recipes, and a recipe is usually one part reference, with its own cameras, models and decisions for the PLC, camera or robot arm. Each recipe has views, one per side of the part, seen by one or more cameras while a robot arm turns it. Each view gets its own dataset and is trained on its own, because defects differ from one side to another.
All of this is created and linked automatically. Parts that pass the line land in the right dataset. In Rely Cloud people label them, by hand or with a model's predictions, train, and the new models show up already linked to their recipes. The user just sees things appear.
To get there I designed for the edge cases: users without the link key, and users who don't want Rely to lead. Rely first can be switched off, and a settings screen maps views to datasets by hand.
The loop closes in production. While reviewing parts the model has already judged, the user picks images to relabel and sends them to that view's dataset, another one or a new one. The labelling happens in Rely Cloud, and they choose whether it syncs back.
2. Engineers build the inspection themselves
We used to write a Python script for each line. I replaced it with a visual editor. The engineer connects blocks for the trigger, the cameras, the detection model and the rule that decides OK or KO, then publishes it to the station. Blocks are named after things they already know, like a camera or a PLC.
3. Showing which recipe is running
As stations and recipes multiplied, people lost track of which recipe was running on each line. A small "live" badge wasn't enough. I replaced it with a production slot: the active recipe plugs into it like a cartridge into a game console.
How I lead
I assigned and monitored work for around 10 interns and two frontend developers. I designed their tests and training, and decided who stayed.
Result
Rely monitors factories worldwide. If it fails, the line stops and the client loses hundreds of thousands of euros. Siali took on the complex projects other companies rejected.