Are great entrepreneurs cut from the same cloth as us ordinary mortals? What psychological traits separate them from the rest of the pack? A lot of VCs and entrepreneurs have pondered about this question, and we turned to DeepSense, a personality profiling tool from Frrole, a Bengaluru startup to check if it had any definitive answers for us. Frrole’s tool is different from the eponymous DeepSense, an AI solutions firm that provides a machine learning lab speeds up the development cycle of machine learning models.
Here’s what we found, after doing a quick survey of some self-made entrepreneurs, bootstrapped and funded, who by and large would be considered successes by any conventional measure. Their attitude and outlook may range from highly pessimistic to highly optimistic, they may be generally unfriendly, with low or average teamwork skills, but the one trait they all index highly on is action-orientedness. A high score on this metric “indicates energy and decisiveness in dealing a tough situation,” reads the explainer on Frrole’s Chrome extension on this metric.
In case you were wondering how you measure up, you can try it the tool out on yourself. All you need to do is install and enter anyone’s email ID, Twitter profile, or phone number on the Chrome extension, or website. In a few seconds, based on a person’s digital footprints from the social and open web, DeepSense builds a personality profile of a person, collating information on a person’s interests, work history, and education. The tool ranks a person’s personality traits on attributes such as attitude and outlook, stability potential, action-orientedness, general behaviour, teamwork skills, need for autonomy, and technical and managerial role fit.
The startup claims that its DeepSense tool, tailored for recruitment and HR teams, is at near-human accuracy in its results when compared to a traditional psychological assessment test. How accurate are their claims, and what’s the underlying technology that helps make these predictions? We reached out to their team to find out.
Founded in 2012, Frrole has gone through several product pivots since its inception — it was an online social newspaper in its early days, and during its angel round in 2014, it was referred to as a big data startup that analysed Twitter data. The startup also retails a product called Scout, a social intelligence tool that provides audience and influencer intelligence, priced between $500 to $1600 per month. On its website, its listed customers include eBay, Flipkart, Grey Group, Smithsonian Institution, and Network 18.
Over an hour-long chat, Frrole Cofounder Amarpreet Kalkat shared his vision for the startup, and the potential applications of DeepSense. Software is still largely dumb, he contends. “This is the age of intelligence, now software is beginning to become intelligent. It looks different from workflow software. Here the goal might be to minimise the effort of the user, it’s a very different way,” he says.
Frrole uses the social web as a data set, to provide consumer intelligence at scale, he says. According to Statista, there were 2.4 billion people worldwide on social media in 2017. “If you can build predictive ability from that data set, you can predict things about everyone, and its useful to pretty much everyone. In recruitment, what if you knew things beyond what a candidate is claiming, for example,” he asks.
Some 63% of employers said one of the top questions they’re trying to answer when looking for candidates is “what are their soft skills?”, he says, citing a 2016 study from CareerBuilder, a recruiting solutions company. “Research proves that if there is an indicator of ultimate employee performance, it is personality and not technical skills,” he says, explaining why they’ve chosen to focus on soft skills over technical skills.
DeepSense helps companies find a candidate with the right mindset for the job opening. “If I was hiring for a financial role, I wouldn’t really mind somebody who is critical, disciplined, not super-friendly even,” Kalkat says. “For a sales role, I might want somebody with very positive attitude, somebody who is friendly, not necessarily critical.”
Frrole’s DeepSense assesses people on the Big Five personality traits, known as OCEAN (Openness, Agreeableness, Extraversion, Conscientiousness, and Neuroticism). “It’s a standard psychoanalysis model (and) probably the most widely used,” Kalkat says, adding that 80% of the Forbes 2000 do psychological assessments for their employees. “We are automating that. Nobody needs to spend 30 minutes and three dollars in our case, you spend 20 cents and a few seconds. All we need is your email ID, and that’s it, here’s the assessment,” he says. Anyone can test out 50 profiles a month on the tool for free by signing up, while paid plans start at $100, for 500 profile searches.
NLP (Natural Language Processing), a branch of artificial intelligence definitely plays a major role in DeepSense, says Kalkat, since they deal with textual data. “Not all of our algorithms are ML algorithms, we use it where it makes sense,” he says. As for explaining the machine learning that goes into it and the accuracy of the predictions made by DeepSense, Kalkat shared a blog he wrote on Monday detailing the science and technology behind it. says that their models borrow from psycholinguistics, behavioural sciences and psychology. It also links to a paper that shows a correlation between personality type and word usage among bloggers.
According to the post, their preliminary studies show an accuracy of 79% to 85%, where test subjects undertook a traditional psychological assessment test and compared those results with the analysis made by DeepSense.
DeepSense only looks into data that exists on public networks, says Kalkat, naming Twitter, LinkedIn, AngelList, GitHub, StackOverflow, as some of the two dozen-odd data sources used for its personality analytics. “We’re not looking into private networks like WhatsApp, Facebook, it’s the public networks, like Twitter which is already open, and it’s a half a billion-dollar business, sharing this data. LinkedIn is partially there, Instagram is there, it will go towards more open, I think. There will be some constraints, but there will be enough data in the digital web, if not the social web,” he says.
“You don’t need much data to make these predictions, by looking at someone’s photo, you could predict with 80% accuracy whether they are straight or gay,” says Kalkat, discussing a Stanford research paper published in February 2017, which stoked privacy fears in the LGBT community. The researchers Yilun Wang and Michal Kosinski had built an algorithm that could use facial images to correctly distinguish between gay and heterosexual men in 81% of cases, and in 74% of cases for women. In his earlier work from 2015, Kosinski along with researchers Wu Youyou and David Stillwell at the University of Cambridge were able to predict a person’s personality on Big Five traits more accurately than a spouse by mining a person’s Facebook likes. The research paper proclaims that computer-based personality judgments are more accurate than those made by humans.
Other notable Indian HR tech startups using AI and machine learning include Belong, an outbound recruitment solution, Param.ai, a provider of screening, retargeting, and analytics solutions, and Skillate, a provider of resume screening software.“ Belong has done more for changing sourcing of candidates. We don’t focus on sourcing, we focus on evaluation and assessment,“ says Kalkat.
“I am aware of one company, Cambridge Analytica, they have this kind of a dataset, they collected it through different means, and at a very large scale in 2013-2014,” says the founder of an Indian AI startup who did not want to be named. Cambridge Analytica’s game-changing impact on US President Donald Trump’s electoral campaign in late 2016 has been widely reported in the past. “My only concern here is the dataset — to make such accurate predictions, it would need a humongous amount of data, which is carefully annotated by trained psychologists. This should be done on a few million profiles, a few million articles, only then I would believe that this is meaningful or correct,” he says.
While Kalkat did not have the exact details on the training data set at hand, he said it was in the range of of 5,000 to 10,000 users. “We did not have a psychologist but the proxy for psychologist has been all these papers published by psychologists, because fundamentally that is the basis of how our tech and algorithms work,” he says.
Thejaswi Udupa, CTO of Apnastock.com, a Bengaluru construction tech company, was sceptical of the tool and the inferences it was drawing, particularly the machine-generated paragraphs that failed on grammar and internal consistency. Take the case of Elon Musk, who scores a 4/10 on attitude and outlook (Slightly Pessimistic), and his personality is stated as “Usually negative and anxious” on the DeepSense tool. Yet the paragraph underneath reads: “Elon Musk is extremely positive and optimistic. He is a usually spontaneous individual and takes life extremely positive and optimistic…”
“Whether the paragraphs are correct can only be accurate as the science has progressed. It’s not even a science — most people would be loathe to call psychology as a science,” says Udupa. “Most people use Twitter to rant about something. Once you do that, it automatically puts your personality as somebody who is negative, or slightly pessimistic. If everyone starts realising that everything I put out in public is being used by recruiters, then you start gaming the system. Therefore, the fidelity of this data right from the beginning is kind of lost, and it’s only going to decrease. How much ever they tweak their algorithms, it’s only going to go down,” he says.
We looked at a few founder profiles together to see if DeepSense was on the ball. Zoho co-founder Sridhar Vembu’s profile, which classifies him as highly pessimistic, with poor teamwork skills did not sit well with Udupa. “If Sridhar is being shown as a 2/5 manager, there is something possibly not as accurate about it. One in ten is highly pessimistic, there is something clearly not accurate about the way this is working. This is because the data it is mining, is not supposed to be mined for this purpose,” he says.
Another gripe we have with the tool is that the DeepSense profile sheds very little light on where it came from — how much of it is being inferred from a profile on Twitter, for example. “We don’t want to share that because its core to the IP, it’s also a function of the thing evolving very fast right now. So we’ve got LinkedIn integration going live next week, which will change these ratios very significantly,” Kalkat says. Assigning a fixed weight to a social media platform is not how machine learning works, especially if it’s the semi-supervised kind, he adds.
At the same time, he says that DeepSense is largely not like a black box algorithm, and a large part of it is explainable. “Maybe 10-20% is a black box, which we cannot explain. Currently, a very less part is unexplainable, moving forward, more will become unexplainable, though you’ve got laws like GDPR coming in, which might require you to share a certain level of detail,” says Kalkat.
We will be trying the best from our side to be transparent, because our principal philosophy is to be as open as you could potentially be, not just with products, but with organisation and all, Kalkat says. A few years ago, Frrole had published details on founder salary and valuation on its blog. It will need a lot more of transparency to convince the sceptics.