We met Gaurav in early 2022 when Raga AI was at the idea stage, an offshoot of his experiences at Nvidia, Texas instruments and Ola. The idea was underpinned by a simple realisation - AI is finding its way into every sphere of life, but continues to be unreliable when deployed in a real-world setting. Gaurav begun building an AI platform to observe deployed models to gauge when they would ‘drift’ or begin to deteriorate in performance. This could happen when the model is deployed in conditions different to what it was trained on, weather for example. As real-world conditions are dynamic, it isn’t surprising that models drift over time. What was surprising though, was this - Top AI teams had insufficient mechanisms to identify which of their 10 (or more) deployed models drifted and took an average of 3-6 months to complete the retraining cycle once they did manage to isolate the issue, which was mostly flagged by a customer who had a poor experience. Raga upends this paradigm with its patent pending embedding generation technology (RagaAI DNA) at its core, which can identify when AI models are drifting in real-time (before it is flagged by a poor user experience). It began working with customers and also set its sights on completing the retraining workflow, by identifying the right data and completing the retraining loop, reducing the entire cycle time by ~70%.
Although users loved their observability solution, educating prospective customers of AI’s post-deployment failures proved to be a challenge. Showed symptoms of being in the new product - new market category (Refer to our Market Centricity Blog). The teams Raga was looking to sell to were from top orgs that followed data science best practices and positioning was key to get in the door. Gaurav and his team were quick to distill the customer feedback and found an alternate way to position the product. What seemed to strike a chord with prospective customers was a pre-deployment testing platform that could help deploy the most robust AI model. Raga, with its drift detection technology was well positioned to pivot to a pre-deployment AI model testing suite
RagaAI DNA: Raga’s Foundation Model:
Raga’s Foundation Model works on proprietary embeddings that map out the model’s training data (what it has learnt from) in an n-dimensional space. With this, it is easy to identify clusters where the model may perform poorly, especially when combined with the 300+ statistical tests that Raga carries out.This is core to finding the deficiencies in data and model training automatically and to top it all, for certain use cases, Raga platform can also generate training data automatically for undertrained areas.
The shift in positioning helped them start closing sales to large Fortune 500 clients and they currently work with 10 customers. We are proud to announce Raga AI’s emergence from stealth with a $4.7m seed round led by us at pi Ventures with participation from global investors including Anorak Ventures, TenOneTen Ventures, Arka Ventures, Mana Ventures, and Exfinity Venture Partners.
What’s next at Raga?
Having achieved the first signs of product market fit is just the beginning of Raga’s journey. Raga plans to double down in the US and target 50 global customers by the end of this year, expanding their team from 40 to 80 people. Competition in the AI testing space is limited and Raga’s comprehensive suite of tests combined with their focus on customer needs position it well on the way to take a sizeable chunk of the growing market