# We're building Continual Learning as a Service

## What is Continual Learning and why is it important?

**For all their incredible feats, LLMs today have a fundamental limitation in comparison to human intelligence: they do not learn from experience.**

You can guide an agent through complex tasks — but like a new employee, it needs oversight. You steer it toward the right solution. The difference? A new hire learns. The next time you run the same task, the agent needs the same feedback all over again.

That's because LLM-based applications only improve in discrete moments — when the base model updates or the developer tweaks the prompt. Both are rare. Frontier providers ship maybe two major updates a year. Developers we surveyed update prompts less than once a month after a project matures. In the meantime, users just have to get accustomed to their limitations

From experience, we know the cost of not improving models quickly. Our team helped build LinkedIn's realtime relevance stack — it resulted in a 20% reduction in dismisses and an increase in retention. Even a one-hour lag in activity features was found to hurt model performance by 3.5%.

Simply put – the lack of continual learning in LLMs is causing users and businesses to churn from applications that rely on them.

## Our mission

**Shrink the time from user feedback to better model behavior — from months to seconds.**

We know that there are already valuable techniques that for all practical purposes can create systems that learn and improve with every interaction. However, building these systems today requires stacking together several capabilities (more info here) that most teams don't have the bandwidth to build or maintain.

Orizu will offer these out of the box to give your customers the continual learning experience they desire.

## **How we get there**

**We have a multi-phased master plan:**

1. **Automate prompt optimization.** Most teams dread this, but it is where feedback turns into improvement, and it's where most teams get stuck today. If you can't close the loop from "user had a bad experience" to "model behaves better," nothing else matters.
2. **Build an active reading system.** Once optimization is running smoothly, the next step is making feedback collection smarter by accumulating knowledge and preferences that also update evals and guardrails accordingly.
3. **Go beyond prompts.** Expand into finetuning, LoRAs, memory layers, and whatever new techniques emerge.

Each step brings us closer to the vision: AI that gets better every hour, every conversation, every turn.

## Join us

## FAQ