Skip to main content

AI/ML in Flipkart--CTO Talk event experience

Our team has done some Machine Learning projects. When it comes to recommendation system, my favorite has been so far PredictionIO which was based on Apache Mahout (Now they have switched to Spark). I was also avidly learning Machine Learning for some time. My boss suggested me to attend the CTO Talk event hosted by OrangeScape. Its first session is “AI/ML in Flipkart” by Flipkart CTO Ravi Garikipati. As I mentioned earlier, I always had my skepticism towards these buzz words.  But, this presentation changed my perspective for good.

Event

The event started with welcome note by Rajesh Manickadas, Head of Engineering, OrangeScape. Then OrangeScape CEO Suresh Sambandam quickly introduced and handed over to Ravi Garikipati.

Flipkart CTO Ravi Garikipati


Ravi explained his journey from TCS in 1988 to IBM, Oracle, etc. He explained his earlier attempts of AI based virtual assistant for [24]7. Also, quickly connected with the audience through his remark about his connection with Chennai and why—his wife and his in-laws.

AI/ML in Flipkart


Ravi then went on to explain the challenges in India and why it needs homegrown MI solution.
  • Flipkart is collecting everything-- from users’ GPS location to page scroll
  • On an average, Flipkart generates 10 TB of data per day; it goes up to 50 TB on sales days like Big Billion Days.
  • Flipkart answer for MI challenge is FDP (Flipkart Data Platform)
  • FDP is built based on Hadoop, Hive, Cassandra, Power BI
  • Flipkart is also investing in GPU based hardware as doing Machine Learning in Laptop will take ages.
  • In India, with respect to data, the challenge is that the whole family may use same account and credit card details. So, personalization is kind of a challenge.
  • Flipkart has open sourced its text processing library called “FastText” (Unfortunately, I can able to find Facebook’s fastText library alone)

Applications

  • From their FDP ML studio, they’re inferring 40+ insights like infographics, brand affinity, fraud, etc
  • Trying to show different homepage for tier-2 places for price and other reasons
  • Trying to show different recommendations for different customers based on price, brand, etc
  • Trying to show trending products and mobile models to mimic real time salesman feedback
  • Trying to understand why they have 30% bounce rate
  • Showing similar dresses based on “features” instead of mere colors
  • In Myntra, they generate rapid designs. Their system comes up with new designs for apparels. They’re sourcing from Tiruppur and Ludhiana; but for designs previously they have to rely on France.
  • Using data to launch their own private label products.
  • Using “contextual NLP” to process sentiments
  • Using NLP and ML for cleaning up reviews (auto titling, fix rating, handle profanity, understand Hinglish)

Ravi’s approach/framework

Ravi suggested that we should always have “AI first vision”. To arrive intuitive features, it is important to know the business.

Audience Interaction

I asked him if they have any plan to open up their FDP to outside. Ravi said they have no plan to launch it as a prediction service. (This is kind of disappointing for me.)

Someone asked if they have any plan to use trust framework like Blockchain. Ravi said they’re using Blockchain in their fintech venture, but not here.

Another person asked why their app is not using AI and he compared it with another Chennai based Startup Mad Street Den. Ravi humbly said he even talked to Ashwini of Mad Street Den and will try to work on improving.

Suresh asked as he heard that in US, Google is panic about Amazon and if similar things are happening with Flipkart. Ravi said, many brands have realized to put ad in the final destination and in Flipkart also they’ve seen huge Ad revenue.

My Personal Take-away

  • Earlier, I was unclear why Flipkart is paying huge salary for its talents and CTO. Now, I realized only companies like Flipkart has the scale to retain such huge talents like Ravi.
  • Ravi’s presentation changed my perspective towards Machine Learning. Ravi’s framework of “AI first thinking” is a good idea.
  • I came to know about INSOFE, Hyderabad. Ravi was sharing good feedback about it. (But, in Quora and Google Reviews, this training institute is having mixed reviews.)
  • Came to know about OrangeScape. I searched and found their product KiSSFLOW. They served coffee and tea, but I’m not sure, why they didn’t promote them. Their office space is quite creative with Tamil inscriptions and interior works. I thought of taking photos and felt that it may not be appropriate without proper permission and so avoided.
  • Also, came to know about Mad Street Den, the Chennai based startup doing AI related works. Probably, it would be a good opportunity to interact with its CEO Ashwini, in another event?
Labels:

Comments

Post a Comment

Popular posts from this blog

Converting PSD with PHP/ImageMagick

After seeing feature rich options in Imagick PECL extension at Mikko Koppanen 's (the author) website and also impressed with ImageMagick 's features, I have decided to use it for the PSD to XHTML conversion website that I'm architecting and managing. Since, the team wants programming help for converting PSD images, I have tried it (documentation is sparse on PSD handling) Converting PSD to PNG/JPEG/etc Note that, flattenImages() is needed for layered/multi-page PSD file. <?php $im = new Imagick('test.psd'); $im->flattenImages(); $im->setImageFormat('png'); $im->writeImage('test.png'); ?> Extracting PSD layers One by one <?php $im = new Imagick('test.psd'); $im->setImageFormat('png'); for ($i = 0, $num_layers = $im->getNumberImages(); $i $im->setImageIndex($i); $im->writeImage('layer' . $i . '.png'); } ?> Note that, there is a better version below In a single call with writeIm...
5 comments
Read more

NETELLER - Privacy and security design flaw

Yesterday, colleague of mine brought to my notice about a payment system called NETELLER and it's merchant API named NETELLER Direct API V4 . This NETELLER Direct API V4, helps the merchant to collect amount from users. It's simple--same just like early Authorize.net , along with the required amount you collect the user's id and password and post them to their API URL and they'll send you back with success or error codes in XML format. The major problem with these types of system is security and privacy --you lose both as you're forced to type your username and password in alien web page. If I remember right, this is was the case with Authorize.net and they changed their design to something like PayPal . The PayPal design is somewhat better as you never type or forced to type your username and password in other alien web pages. The alien merchant web page uses NETELLER Direct API V4 is forced to get user's NETELLER account and password. I'm much s...
Post a Comment
Read more

Stampede and the "Dirty" "Dark" Crowd

Actor Ajith Kumar’s recent interview has sparked quite a few conversations on social media. The part that caught my attention, was his take on crowds. About 30 years ago, when I first joined a college in Madurai after growing up in other places, I experienced a few cultural shocks. Perhaps these weren’t unique to Madurai, but that’s where I first noticed them. One major thing that stood out was the behavior of crowds. For instance, if you suddenly see people rushing to board a bus, chances are there’s a pickpocket in action. During one of Madurai's annual festivals, I noticed some young men carrying water bags — not to distribute water, but to spill it on women, often on their chests. Shockingly, this was almost normalized; parents would quietly tolerate it to avoid public embarrassment, walking a little farther behind the crowd. When some women happened to witness this, they would just shoo the boys away instead of confronting them. The crowd, in such cases, became a kind of...
Post a Comment
Read more