How hila wins with Reinforcement Learning

When an established BI vendor sent in a swat team to deploy generative AI on a customer’s data, they spent nine weeks of heavy engineering and still landed at only 65% accuracy.

The same customer tested hila, and hila hit 95% accuracy in just two weeks — results confirmed by an independent auditor. With continued fine tuning over another two weeks, hila pushed past 99% accuracy.

At the heart of hila is Reinforcement Learning (RL) — the reason hila achieves accuracy and speed that others can’t touch.

Why Reinforcement Learning (RL) Matters

hila doesn’t just generate outputs - it learns and gets better with every interaction. RL allows hila to capture three layers of knowledge that LLMs always struggle to integrate:

This layered knowledge modeling, continuously refined by RL, explains why hila can improve responses coming from the LLMs.

Example, in a previously published blog, we detailed how a foundational LLM is only correct 12 percent of the time, but with hila’s additional reinforcement learning layers, this accuracy increased to 100 percent.  

How RL Works in Practice

hila provides several components for implementing this RL flow:

• Feedback — Users can “thumbs up” or “thumbs down” an answer.

To act on this user feedback, hila supports additional RL tooling, including:  

• Jargons — If there is a disconnect between usage of common terms among users, hila RL is used to teach the system. For example, one customer has the vowel “e” that means net profit. With hila, they can ask, “What was e by product in q2 2025?” And receive comprehensive analysis.  

• Custom Questions — In other situations, hila provides RL tooling for more complex questions. Through “custom questions,” formulas can be defined in natural language. The system then generates examples of responses against the company data and allows users to provide feedback and propose changes. Specific terms can be defined, and along with them, associated formulas to enhance the responses (e.g. What is the quarterly phasing for Tires for FCx 2025?).

With jargon and custom questions, every piece of feedback becomes part of the reinforcement learning cycle.

Over time, hila gets precise and tuned to the working of the enterprise.

The Impact of RL  

Because RL powers the entire system, hila eliminates the expensive services and rigid BI software layers that have built up around enterprise data. What used to take five business days and $20,000 for a new report, now takes seconds. There is no need to directly involve technical users - the domain experts can provide this RL feedback directly.  

This is more than mere efficiency. It’s a new way of working — where RL ensures that answers keep improving in real time.

That’s the power of hila.

See it for yourself — request a demo today.