At ML/AI@Scale by Freshworks, lessons in imagination

A primary motivation for artificial intelligence is to reimagine for an automation-first world. 

A half-decade ago, a customer would call a company or email it seeking a solution to some problem. A support executive would then forward the matter to the relevant team to troubleshoot. Now, customers are on smartphones and want immediate fixes… Else, there’s Twitter.

How do you reimagine for this world? How do you build intelligent models to understand and respond to customers or users spontaneously, or have meaningful conversations with them? 

Artificial intelligence (AI) and machine learning (ML) technologies might seem the obvious answer. 

But if that were the panacea, we wouldn’t be going all-caps on John the bot or cursing at that careers app for another vexing recommendation. Ms. Vice-president of Sales wouldn’t be blindsided for a lack of context in chasing leads.

The AI/ML world is vast with a lot yet unexplored. A few companies, though, are pushing their engineering chops to get that bot, that jobs search engine, and the sales dashboard to understand exact requirements, extract appropriate data, and pull the right context for delivering perfect solutions.

There’s much to learn from them.

At Freshworks’ fifth flagship engineering mixer, Saas@Scale, a handful of experts came together to illustrate how they solve for the future using machine learning and artificial intelligence. The previous editions had dwelled on data, security, DevOps, and platforms.

“We would like to establish India as a product nation. And that’s the purpose of Saas@Scale,” said STS Prasad, the senior vice president of Freshworks, inaugurating the event in India’s tech hub Bangalore. “We are looking at Saas@Scale as that forum, that community where technology experts can come together to share their experiences, and gain from each others’ experiences.”

Read on for snapshots from ML/AI@Scale

How LinkedIn powers discovery for job seekers

The world’s leading professional networking platform has thousands of members seeking to connect with other professionals, search for training and learning programs, or hunt for better career opportunities. This has LinkedIn constantly innovating to enhance its discovery, search, and recommendation engines. “The mission is to connect talent with opportunities. And the opportunities come in different scales and varieties of times,” said Nagaraj Kota, senior staff machine learning engineer, LinkedIn AI. 

That’s over 20 million dynamic job postings with limited lifespans. Nagaraj dived into a technical talk on understanding job-seeking behavior and guided the audience through how LinkedIn brings together elements such as search, alerts, navigation, and recommendation for optimized results. 

“What is the meta level problem you want to solve? If you are in the job market, you don’t know what you want to search for. People run queries based on skills but there is a lot more we can do to help them to understand… this is the right query that you have to do to get to the right job that you want to do.”

Reimagining bots at Freshworks

Software-as-a-service provider Freshworks has customers across the world who speak in more than 30 languages. But let’s just consider English. How do you make a bot understand language, intent, and context, and respond appropriately? How does a bot fetch the right answer for the same question asked in different ways? Or identify small talk?

Tarkeshwar Thakur, vice president of engineering at Freshworks, immersed the participants in a session on how the company’s answer bot is built to effectively solve for such issues. His talk included a generous sharing of information on the tools used to, say, detect small talk, incorrect spellings, or even gibberish.

A critical part of this is teaching a computer or bot what a word really means. 

“If you want to understand what the meaning of a particular word is, you could look at what context that word appears in across a large collection of documents,” said Tarkeshwar. “Intelligence comes from when you pack a lot of the little things right. You start to get a feeling that, oh, this thing is intelligent.”

Bigbasket’s 6 R’s for analytics at scale

India’s largest online grocer nearly tripled its sales volume in the last year, operating out of about 100 facilities across 26 cities. How did it do this? Largely by developing an analytics engine to alleviate the drudgery of grocery shopping. To grow at this scale, a primary ask of this engine is that it needs to be robust. “This has got to be something that operates at exponentially growing volumes,” said Subramanian M S, head of analytics at Bigbasket.

The analytical engine also needs to be adequately responsive to handle thousands of requests in milliseconds, rigorous, reliable 24×365, relevant in improving key business metrics, and, finally, generate returns in multiples on the investments made.

This big challenge in all this is forecasting. “A very smart person many years ago said this very correctly, that all forecasts are wrong,” Subramanian said. “Just that all of us are working on reducing that error in the forecasts.”

Customer engagement, the Freshworks way

Let’s begin with the scale that the Freshworks AI/ML platform has to optimize for—more than 200,000 businesses across 145 countries—which makes multitenancy one of the biggest challenges before this platform. “Every element, feature or capability that we build should work for 200,000 businesses (and their customers),” said Swaminathan Padmanabhan, director of data science at Freshworks. “We are talking about billions of people.”

To solve for this, Freshworks has structured its investments in its AI platform as six layers, each with sophisticated capabilities on top of the one beneath it. At the bottom is the data representation layer that processes and interprets text and voice inputs. On top of that is the knowledge layer that collates data signals from Layer 0 and forms blocks of knowledge. The layer above identifies what blocks are relevant for a particular business scenario, user persona, or business metric. And above that are the forecasting, recommendation, and autopilot layers.

“By structuring our investments this way,” Swaminathan said, “it becomes easier for us to roll out and manage new initiatives.”

Rethinking AI/ML culture from xto10x

Culture plays as important a role as technology in building powerful AI/ML platforms. But often, startups get it wrong. “The real challenges are not in the models or the tools, but beyond that,” said Goda Ramkumar, general manager, datascience, at xto10x Technologies. She traced a typical journey of a startup before suggesting tweaks that could set companies on the right path.

This is that typical journey: collect all the data a startup possibly can, hire data scientists, figure what they can solve for with all the data collected, and, finally, the litmus test to gauge if the effort has an impact.

The first stage, on the other hand, should be for startups to define their business goals and then decide what they want data to achieve for them, Goda explained. Stage 2 should be for data engineers to make all that information accessible and usable before data scientists can identify what problems they can solve with the data collected.

As for the last stage—the impact test—here’s Goda’s recommendation: “Instead of doing everything and then asking if it had an impact, the first questions should be how will we measure it and what is the right experiment designed for it?” 

For more such engineering mixers from Freshworks, visit Saas@Scale