Coverage by Bhat Dittakavi of Variance.AI on “Myntra” with Jeyandran Venugopal, CTO of Myntra at CIE@IIITH on 20th July 2017
At Myntra, we are in the business of Price Elasticity of demand slightly in our favor and consumer is not that sensitive to price as we are in fashion unlike regular retail. Choice Theory as well. We built first-of-a-kind Rapid Design system using which we release designs inside 15 days. We use generative discriminative networks of machine learning to generate new designs. Rapid acquisition of users is one asset. We incentivise to retain customers and do targeted promotion.
Size Fit & Merchandising
We are looking at data of the previous purchases. We create clusters around it. We try and match brand preferences within a cluster. We rolled this beta and we are seeing really good results. Now that we know the brand preferences, we can have better predictions on the size that suits our repeat customers. This is design-based prediction. Though we have US and UK sizes for cardinal dimensions (like shoulder width), they may not work as they won’t work with the height as it may show off the midriff!
Least undesirable tradeoff to take? Shoulder size, length or loose or tight fit? This is not easy problem to fix. We classified users into different body size segments. Boxie, lean and tall and so on. If a segment picks this size of this brand, we calculate the likelihood of the return and we do better recommendation. There is a cost to returns such re-evaluation, cleaning, QC and re-stocking.
Q1 Returns: User never used and the returns came as original with tags and all. Each mobile renders the product differently on its display or the feel of the fabric can’t be easily communicated. These are common reasons for returns as purchase thought one color or texture and reality was different.
Q2 Returns: Tried and returned. This goes through hygiene refurbishing process.
Defect Returns: Items with manufacturing defects will be sent to the supplier.
Others: Liquidation sale to vendors who buy in bulk.
We got ways to figure out the body shape of the user using depth sensors. Even if you have the 3D model of the user, you have to create the 3D mesh model of the fabric. Fabrics are stretchable and the demographic and intended fit of the fabric is different and we got to take note of these nuances. If you let the user wear it digitally, when he sees himself in the dress, the desire of purchase may go down as common users don’t look like models. We have to find the right balance! Fashion is aspirational and appeal-driven.
We are looking to outsource our catalogue operation. We do cataloguing operation both in Bangalore and Delhi. We have in-house models too. Granular segmentation of the user works well for us.
Handling the Returns
Automated QC on returns using CV. Can’t train the labor to catch the defects in our products. Customers return a leather item. By design leather products have wrinkles. When customers looks at the leather, he may think it is damaged whereas it is the nature of the beast. We leverage computer vision on both forwarding and return channels.
Q) Search domain in fashion?
Text search (though the corpus is limited) gets harder if the user comes for window shopping! We need to figure the intent before showing the results. This is a challenge. Can’t afford to show the use results that won’t match his or her intent.
In the default page, we show 48 articles. When you do something and come back, it shows different 48 articles based on your click through pattern using clustering. We refine search results as the click through happens.
Q) Logistics after the purchase? Do you do any last mile optimization?
Alternate delivery models. Local kirana receiving it and then dispatching! Last mile delivery is a challenge. We use past history of delivery, we learn from it the time it takes to deliver. We use it when we promise to the customer. CPD (Customer Promise Date) breach is critical and we want it less than 1%.
Twice the day delivery teams go for milk run both forward and backward. We leverage user’s past undeliverable data to slot the delivery times. I personally did 15 deliveries to understand the last mile. These delivery guys know the local areas very well and then we have huge levels of repeat buyers and hence the delivery boy knows them very well.
Q) How to handle churning workforce?
Q) Customer care?
Interaction to order and cost for interaction are two big metrics. We want to use clear answer to the customer. No one likes IVR. We got conversational bots. We haven’t done any Indian languages. We track customer satisfaction through NPA score. We send the survey to the customers after the call. Though small percentage respond, that sampling is good enough through statistics. We try to send the survey as quickly as we can. Improving NPA score while not incurring huge cost is good.
Bot hands off to real agent when it bumps into a problem. Many a times, it is about where is my order? guy never came and delivery attempted message is shown and so on.
SDA (Service Delivery Agent) app for delivery guys. When he delivers, the app stores the GPS location and also comments.
Each article on clothing, there is a one inch tolerance on say shoulder width. Our vendor won’t accept returns on less than one inch deviation. When the customer orders it on tighter side and the a 40 inch shoulder shirt got only 39 inches, return will follow.
Q) Do you have any way to capture the user’s all measurements in one go?
In mass customization market, every piece of information extra you ask the user, it drops the conversion 50%.
We figure out a way to price using price elasticity and machine learning. We do differential pricing and do targeted coupons.
Q What technologies do you use?
Supervised, Unsupervised, Clustering
R, Python, Tensor Flow and much more.
High Revenue Day (HRD) or End of Season events where we do TV and news paper advertising. Otherwise we do digital advertising.
Omni-channel Store: We have a brand experience tool in Bangalore where we experiment with VR and other technologies. We do engagement drives in store. Automatic checkout! Virtual wardrobes! Use gestures to navigate the catalogues. Use more engaging and light weight presentation of the articles. We do A/B testing also there. Off-line is for more of experience and brand channel. Offline retail people see a breakeven in 3 years. We will breakeven this new store in 5 years. We are doing reasonably well. We look at virtual trial rooms in the store.
We have now become brand partner for Mango and we are setting up physical stores entirely using all of our technologies. We also manage their digital platforms. We give our ecommerce in a box to other brands.
$25B of $100B Indian apparel market is branded fashion. We see bigger pie in grabbing the offline numbers. We are doing half of online apparel sales. We can go after the long tail go after offline customers and bring them online.
We segment users differently.
C1: Existing Myntra online users
C2: Existing non-myntra online users
C3: online and not fashion buyers
C4: Never bought online
We are coming up with an open challenge and a good winning prize to announce soon. Any team in the world can start working on it and start submitting solutions.
Q) Have you cracked the unit economics?
We are profitable at Contribution margin. If you add G&A and others, we are getting EBITDA to zero now. Myntra is run as an independent company.