About the lab
A small lab with a narrow job: make AI useful at work.
Rilibay started in 2024 as a software studio with one rule. Build focused tools that do one job well, and charge a fair price for them.
Our first product, Rilifi, applied that rule to live audiences. It now runs quizzes, polls and spin wheels for creators, trainers and event teams in more than ten countries. Rilifi taught us what it takes to keep software working for real people. It’s dull work: uptime, support and small fixes, week after week. It matters more than anything else.
Now we bring the same habits to AI. Models improve faster than most businesses can keep up. The model is seldom the problem. The problem is everything around it: the workflow, the data, the odd cases, the review step and the screen people use. That’s the work of an applied AI lab. It’s the work we want to do.
01 - Philosophy
Why practical AI, not hype.
Most businesses don’t need an AI strategy. They need back the three hours a day their team loses to copying, checking and reformatting.
That’s less exciting than a demo. It’s also far more useful. It means we pick narrow problems. We measure results on real examples. We keep people in charge. And we say “this isn’t a job for AI” when it isn’t.
02 - Standards
Engineering standards.
The rules every build follows, whether it is ours or a partner’s.
- 01
Evaluation before release
We test every AI feature on a set of real examples we have permission to use. When the model or prompt changes, we test again.
- 02
Human review by default
Tools draft and suggest. People approve. We let a tool act alone only when the evidence says it’s safe.
- 03
Least-privilege data access
Each tool sees only the data its job needs. We log every access and mask sensitive fields where we can.
- 04
Model-agnostic by design
We aren’t tied to one AI provider. We pick a model for each task based on tested quality, cost and where the data must stay.
- 05
Observable in production
We track errors, speed, cost and output quality. That way we catch problems before users report them.
- 06
Fast, accessible interfaces
We design for WCAG 2.2 AA. Pages load fast on ordinary phones and networks, not just the latest laptop.
03 - Stack
What we build with.
We choose tools per job and change them when the evidence says so. This is the default kit.
| Layer | Default choice | |
|---|---|---|
| Interfaces | TypeScript, Angular, server-side prerendering | Fast pages that work for every person and every search engine. |
| Infrastructure | Google Cloud and Firebase | Secure, managed parts, so we spend our time on the product. |
| Real-time | WebSockets | Live sessions, like Rilifi’s, need updates in under a second for big rooms. |
| Models | Frontier models from several providers; open-weight models where data must stay in-house | We pick the model for each task by test results, not by brand. |
| Evaluation | Task-specific test sets and regression checks on every change | Quality you can see in numbers, not adjectives. |
Want the long version? We write about our methods in Lab Notes.
Start a conversation
Bring us a workflow. We’ll tell you if AI should touch it.
Thirty minutes, no slides needed. You describe the work. We give you a straight answer on whether an AI tool would help, and what it would take.