The Story of Freddy & the AI/ML architecture at Freshworks
Like Suresh Sambandam, STS Prasad also has a deep love and fond memories working with early products of RDBMS like Sybase & Oracle. This blog is an excerpt from the talk (wealth of knowledge) by STS Prasad for CTOTalk where he shares about the analytics platform and Freddy the AI service of Freshworks.
Freddy and Analytics platform — How it began?
Investments began to create “Freddy” in 2017 to provide AI/ML capability across all the products of Freshworks to benefit users.
In 2018, Freddy was introduced for help-desk executives to handle the support tickets effectively via automatic reply, auto-detect the priority, etc. In the same year, embedded analytics from within the product was moved out into providing an analytics platform as a service across all the products for a seamless reporting experience. Following that, Freddy was introduced for Freshsales in 2019 and for Freshservice in 2020.
Freddy is very specific for each product. i.e. Freddy for Freshdesk is unique and is not shared with other products like Freshservice or Freshsales.
Freddy architecture
- Primary consumers of Freddy are actually the users of Freshworks’ products. Freddy is designed with a vision of “AI+Human” and not “AI vs Human”.
- The core data foundation for Freddy is laid on BAIKAL which is a Hadoop infrastructure. Data needed by Freddy is made available in BAIKAL in around 4 hours time period, so there is no load upon the production database by Freddy.
- There are 2 types of ML models —> account-specific models (data of a specific customer) and global models. E.g. In Fresh success, to track the churn risk of the customer, specific customer characteristics are derived to segment customers.
The challenges in bringing AI into products
- Make the product team and engineering team think with an AI mindset.
- Domain experts with AI expertise — The combined talent is hard to find.
- Freddy team embedded in each product, that completely focuses on AI-first thinking within the product. There is nothing like a common Freddy team across products.
Questions by audience and reply by STS (related to Analytics)
Some of the interesting questions and the responses from STS Prasad were:
When was the first data warehouse set up for Freshworks dedicated to OLAP?
“Sometime around 2014, the separation of OLTP db and analytics db were done for each product. Amazon RDS was used for the production database and Amazon Redshift for data warehouse needs within each product. This change was done very early in the product cycle. It needs to be done fairly soon because the reporting workloads need not be loaded over the production database. If the same db is used for OLTP and OLAP needs for a too long period, it would pivot the transaction capabilities of customers.”
Any reason behind choosing Snowflake over Redshift?
“Performance of queries is much better in Snowflake. Certain housekeeping jobs which are needed in Redshift to maintain performance and storage are automatically done by Snowflake…especially complex queries.”
Are AI features rolled out to a small set of users? How are they monitored?
“The rollout process for AI features will be much slower. It is made available to all customers after multiple iterations of beta checks, maturity level, and usage of that AI feature. Also, some AI features will be restricted to certain customers. For example, a feature that segregates tickets into high/medium/low makes sense only for those customers with more data. Very small data doesn’t make sense in training that account-specific model”.
Can external people contribute ML models to Freddy?
“That's a good idea….but not available at this point of time.”
Overall, it was quite an insightful session with STS Prasad at CTOTalk’s monthly deep dive. A Saturday well spent.
