Building advanced, in-house NLU without an extensive research team

The Challenge

Meekan, now part of Doodle, uses natural language understanding to interpret users’ requests when scheduling meetings in Slack. Initially, the company relied on LUIS, a cloud API by Microsoft, for extracting intents and entities. However, this approach created two major challenges for the growing company:

• No potential for customization, to improve performance in Meekan’s specific domain, scheduling
• No ‘ownership’ of their AI, depriving them of a strategic advantage over their competitors. Furthermore, the data remains with Microsoft, improving their models.

Meekan wanted to take their NLU in-house, but didn’t have the resources to build LUIS from scratch.

The Solution

After evaluating different solutions, Meekan adopted Rasa’s toolkit. Rasa gave Meekan the ability to customize and take full control of their NLU without a big research team.

Using the fast migration guide, Meekan went from idea to production in four weeks, in part by importing existing training data directly from LUIS. After a few tweaks and internal tests, the NLU model was ready, and the switch to Rasa was completed without interruption for end users.

The Results

Meekan overcame the challenges associated with LUIS by successfully implementing Rasa:

• Achieving better NLU results by tailoring underlying algorithms to their domain
• Establishing a competitive advantage by owning their training data and underlying technology

Meekan also decreased the latency period, meaning faster response times at scale.

We recently switched from LUIS (Microsoft) to Rasa to be in full control of our AI: we can tweak the performance, we know where the data is going, and we can scale up without worrying about surprise bills

Eyal Yavor, CTO Meekan