(From larboard to right) Rapid technology project’s Yash Kotak and Rohit Chauhan with arch artefact administrator Ambarish Kenghe at Myntra’s appointment in Bengaluru. Photo: Hemant Mishra/Mint
Bengaluru: In October 2016, Yash Kotak, Rohit Chauhan and added associates of online appearance banker Myntra’s aggressive Rapid technology activity hit a roadblock. The aim of the activity was to bear fast appearance products—fast appearance is a appellation acclimated by retailers to alarm the dispatch up of assembly processes to get new trends to the bazaar as bound and cheaply as possible.
Essentially, the plan would future-proof its business—understand demand, acknowledge bound to new appearance trends, cut costs, and abate discounting in a fast and cost-effective manner. To apprehend these goals, Myntra started experimenting with bogus intelligence (AI) systems that accustomed shapes, patterns and colours to aftermath accoutrement that met accepted appeal at a acceleration that would be absurd for acceptable accoutrement makers to match.
Initially, Myntra’s plan was to analyze underserved spaces, bound accomplish baby batches of articles according to the latest trends in those spaces and advertise fast.
The Rapid tech belvedere analysed sales abstracts from Myntra and Flipkart websites (Myntra alloyed with Flipkart in 2014), as able-bodied as the latest appearance trends calm by canoeing Instagram, Pinterest, appearance magazines and agnate sources. Rapid’s technology would appear up with a account of attributes: types of collars, sleeves, cuts, colours and added appearance currently accepted with customers. Here was appearance via engineering.
The activity was accomplishing analytic well. Based on the platform’s analysis, Myntra’s appearance designers had created a cast alleged Moda Rapido (Spanish for fast fashion). But this adjustment was apathetic by the standards of the tech world. The apparatus would bandy up suggestions in argument anatomy afterwards crunching massive amounts of data, a action that took weeks. Moda Rapido, which was aboriginal alien by Myntra in September 2015, was bringing in sales of Rs2-3 crore every month, but wasn’t ascent fast, not by the standards of tech.
Only a few months afore this, Myntra’s arch controlling administrator (CEO) Ananth Narayanan was blame the Rapid aggregation to agreement as berserk as they wanted. Rapid was one of the moonshot bets at Myntra and it didn’t bulk the aggregation actual abundant either. At the time, it alone had 20 full-time staff, artlessly led by Kotak, a 25-year-old artefact able and above entrepreneur, and 47-year-old Chauhan, accession artefact able and above entrepreneur.
Start to finish
In October aftermost year, the Rapid aggregation started alive on an abstraction that had been explored afore afterwards success: what if you automate the appearance architectonics action from alpha to end? If machines could acclaim accouterment attributes that barter currently favoured, shouldn’t they be able to amalgamate those attributes into accepted final designs? The Rapid aggregation started with T-shirts, because of their about simplicity.
But the engineers knew the aforementioned admission would not work. Instead of allurement their computers to beforehand T-shirt attributes in text, the Rapid aggregation wrote algorithms that would get their machines to aftermath images of T-shirts based on the apparent favourable attributes.
Only two years afterwards it started the Rapid technology project, Myntra had auspiciously launched articles that had been angry about in beneath than 35 days, masterminded by machines and engineers who knew actual little about fashion.
This did not assignment at first. The Rapid computers were clumsy to get a bright picture—quite literally—of the images they were crunching, so their achievement was abysmal. The images actuality fed to the machines bare to be of a far college resolution. Back the engineers added the resolution of the images, they begin their machines didn’t accept the accretion adeptness appropriate to action such abundant images. Kotak and his aggregation purchased graphical processing units (GPUs), ultra-powerful machines that are acclimated in AI-related work.
The new machines started bearing designs. The aboriginal ones were still hideous. The engineers tweaked the algorithms and, a few iterations later, there was improvement. They again the process: actualize the algorithm, accumulate agriculture angel abstracts to the GPU, vet the new design, abuse the algorithm accordingly.
In January, they anticipation they had a set of some 30 presentable T-shirt designs, created wholly by the Rapid tech platform.
“Before October (2016), we were calling out aggregate by argument because we capital to accord admonition to the designers,” said Kotak. “When you appetite to alarm things out you accept to analyze them. But back you use images you don’t accept to alarm things out. The apparatus can accept the images in an abstruse way, accept the attributes by itself and absorb them into the design.”
The Rapid aggregation assertive their administration to booty the machine-designed T-shirts to the market. In beneath than bristles weeks the articles were ready. They went alive at the end of February. These absolutely automatic designs were created alike as Moda Rapido designers affiliated to accomplish articles based on the recommendations by the Rapid technology. The catechism was: which would advertise better, the machine-augmented designs or the machine-created ones?
“We debated whether we should tag them as machine-generated, but we absitively not to,” said Ambarish Kenghe, arch artefact administrator at Myntra. “Only one artefact administrator and I were acquainted which designs were machine-generated. We didn’t appetite Myntra advisers affairs aggregate aloof because they were aflame about the machine-generated ones.”
It angry out that one class of the machine-created designs outsold the commensurable class in anniversary of the added 12 clandestine brands endemic by Myntra. Some of the added machine-created designs awash poorly, but the aggregation knew they were assimilate article absolutely big.
By this time Myntra had added categories such as trousers and kurtas to the Moda Rapido catalogue. It additionally started alive on accession fast-fashion brand, Here&Now. Here&Now offers agnate articles to Moda Rapido, alone hardly cheaper. Best Here&Now articles advertise in the bulk ambit of Rs400-1,200. As the technology abaft Rapid improved, both Moda Rapido and Here&Now began to action added circuitous designs, such as printed kurtas and T-shirts, broken jeans and arrested shirts. There’s currently a absolute to the array of designs these brands can produce—the technology hasn’t able a point area it can aftermath things like T-shirts with circuitous cartoon or shoes or accessories. In addition, back it is alive all by itself the belvedere can alone architectonics T-shirts; all the added types of accouterment are still put calm by designers with the advice of machines. But Rapid is convalescent fast.
In the pink
By September, Moda Rapido and Here&Now were calm bringing in sales of Rs12-13 crore every month, authoritative them amid the top 15 brands at Myntra. The two accord 12.5% of Myntra’s clandestine brands sales, said Narayanan. Moda Rapido and Here&Now are Myntra’s best assisting brands, breeding net margins that are about 10 allotment credibility college than the others, he added.
The Rapid aggregation able the eyes of Gautam Kotamraju and Ganesh Subramanian, two above arch admiral who started the Rapid activity in aboriginal 2015 afore abrogation the aggregation beneath than a year later. Kotamraju and Subramanian— the aboriginal is a appearance artist and the additional a appearance veteran—wanted to actualize a fast-fashion cast that was data-driven. They approved to get articles in customers’ calmly aural 45 canicule of sending the final designs to the branch (this action about takes added than 180 days). They additionally capital to see if appearance designers could be wholly replaced by machines.
What we’re accomplishing is not aloof a technology break or a accumulation alternation solve. It’s about abolition the way brands are built.- Rajesh Narkar, vice-president of Myntra’s clandestine labels business.
Only two years afterwards they started the project, Myntra had auspiciously launched articles that had been angry about in beneath than 35 days, masterminded by machines and engineers who knew actual little about fashion.
“The bulk of the Rapid abstraction was to do fast appearance in an able way, accustomed that we accept both a tech and a appearance DNA,” said Kenghe. “In the antecedent appearance of the project, there was beneath apparatus and added designer-input. Over a aeon of time, there’s been added apparatus and beneath supervision.”
Creeping to creative
Fashion is hardly the alone artistic acreage area technology is actuality acclimated to transform the architectonics process. Apparatus acquirements and earlier technologies are actuality acclimated to capricious degrees in fields such as architecture, car design, website architectonics and alike in authoritative music. For instance, an AI-based platform, Shimon, developed at the Georgia Institute of Technology in the US, is already basic new music. Platforms such as The Grid and Firedrop are automating website construction. The declared cold of Google’s Magenta activity is: “Can we use apparatus acquirements to actualize acute art and music? If so, how? If not, why not?”
In the apple of appearance design, several companies and start-ups accept approved to automate design, decidedly in the aftermost year, as AI technology improves by leaps and bounds. Some appearance brands are experimenting with IBM’s supercomputer Watson and Google’s technology accoutrement to actualize designs. StitchFix, an American online retailer, has a tech belvedere alleged Hybrid Designs agnate to Myntra’s Rapid that automates appearance architectonics application AI. Dozens of AI start-ups in fashion, including Stylumia, launched by Subramanian, the aforementioned Myntra controlling who helped alpha Rapid, are attempting to do what Myntra is accomplishing and more.
But Myntra has gone added than the others in applying technology in appearance design: it seems to accept ample out a solid process. Broadly, there are three affidavit why Rapid is advanced of added agnate efforts. Myntra, which was started in 2007, has lots and lots of sales and browsing data. Afar from its own platform, the aggregation has admission to chump abstracts on Flipkart and Jabong. Calm these three platforms ascendancy almost 70% of all online appearance sales in India, according to Myntra estimates. Two, Myntra has the banknote to beforehand in AI technologies, such as GPUs, and allure top engineers. Three, it has congenital adeptness in the accumulation alternation ancillary of appearance and has abutting relationships with suppliers and manufacturers because of its ample clandestine characterization business, launched in 2012. Myntra is able to argue suppliers to compress the accomplishment action from 180 canicule to beneath than 45 for its fast- appearance products. Abounding companies could accomplish one or two aspects but few can amalgamate all the three.
“What we’re accomplishing is not aloof a technology break or a accumulation alternation solve,” said Rajesh Narkar, vice-president of Myntra’s clandestine labels business. “It’s about abolition the way brands are built. Today, we’ve bargain the (manufacturing) action to beneath than 45 days. The abstraction is to do it alike faster. We’re allurement ourselves: if there’s chump absorbed to buy something, can we serve it then?”
“Lots of acceptable abstracts with characterless (algorithmic) models will do bigger than abundant models with bottom data,” said Chauhan. “That’s the accepted actuality of AI. No one abroad has the bulk of abstracts we do. Let’s say added start-ups body models that are bigger than ours. Do they accept abundant cartage to monetize that? The millions of users we accept advice a lot. And the third allotment is, how fast can they iterate on the accumulation chain.
Manohar Kamath, arch of Myntra’s clandestine characterization business, says that ‘no artist job is activity away’ because of machines designing clothes. ‘But I will accomplish them super-efficient. So it’s a admixture of 90% arrangement and 10% artist input,’ he says.
Yet, conceivably best acute is the abiding commitment. Myntra has been alive on the Rapid activity for about three years. Its alternation is assuredly actuality rewarded. Accustomed the attributes of technology and AI, it’s acceptable added companies will amount out automation in design, but Myntra has an important headstart over its rivals.
“The aftermost accession in appearance was Zara with its accumulation alternation model,” said Narkar. “The abutting innovations will be led by AI. We’re aloof off the clue first.”
Top to bottom
Apart from the banking impact, Rapid’s technology is transforming the way Myntra works. The aggregation has absitively that all new clandestine brands will be created application Rapid. Admitting Myntra can’t do abroad with designers, the actuality that all of its new brands will be machine-augmented to a ample amount is a abolitionist change in itself. The aggregation is exploring ablution brands in lingerie, watches and home capacity based on Rapid.
Myntra is application the Rapid belvedere in added genitalia of its business afar too. Artefact descriptions on its site, appropriately far accounting absolutely by agreeable writers, are gradually actuality automatic with the advice of Rapid. It has created a Trends Store on its app that curates the latest trending articles and shows them alone to fashion-conscious customers. In addition, the aggregation buys articles from third-party brands based partly on the recommendations of the Rapid platform.
“Now, we accelerate an architect and a client to go to brands (for sourcing products),” said CEO Narayanan. “Many of our cast ally accelerate us a account of articles in advance. The architect crunches the abstracts to amount out the anticipation of how abundant what will sell. We use that to adjudge how abundant and what kinds of artefact we should buy.”
Myntra now affairs to alpha affairs the software belvedere as a service. It is in talks with added than bristles retailers and brands to affirmation up to the platform. It beneath to name the retailers but said that the partnerships will be able by the end of January. The aggregation has appointed Anurag Asthana, vice-president of artefact development, as the project’s business head. Aural the abutting six months, Myntra affairs to aggrandize the Rapid aggregation to 60 bodies from 35, bisected of whom will be technology and artefact experts.
Next big thing
While Myntra has the bend over rivals in architectonics automation, it’s hardly an insurmountable lead. The aggregation uses a aggregate of open-source accoutrement such as Google’s TensorFlow and Caffe and again builds its own technology on top. Added start-ups can do the same. Myntra’s attribute-extraction adjustment to automate architectonics works, but it’s not article that scales up actual fast. Its adjustment has limitations in agreement of the complication of appearance it can design—remember, currently it can alone actualize T-shirts by itself. The attribute-extraction adjustment is additionally not calmly replicable beyond categories. For instance, admitting Myntra auspiciously created T-shirt designs, it can’t use the aforementioned action for jeans or dresses. From a abstruse standpoint, this a astern method: it favours abstracts analytics over apparatus learning, which is currently the best broadly acclimated anatomy of AI. So it’s no abruptness that the company’s engineers are consistently experimenting with new technologies.
In October 2016, back the Rapid aggregation started exploring the abstraction of breeding absolutely automatic designs, some of its engineers came beyond a technology alleged GANs (generative adversarial networks). GANs were invented by Ian Goodfellow, a researcher at Google.
Neural networks, which comprise affiliated transistors that carbon the anatomy of neurons in the animal brain, are a accoutrement in AI. GANs go further. They comprise two neural networks. In Myntra’s case, for instance, one arrangement comes up with the designs afterwards actuality fed data. The added arrangement alarmingly vets the designs and gives acknowledgment to the aboriginal arrangement on area it’s activity wrong. Through this interplay, the two neural networks are able to anticipate like humans.
Inside Myntra, the Rapid activity has created ramifications that are not anon visible. One of the affidavit the aggregation is acknowledged is that it has conveyed an angel of actuality a appearance destination to customers.
For about one year, Myntra’s Rapid aggregation able and tweaked the GANs. They consulted with Goodfellow on how best to use this beginning technology. Finally, in September, the Rapid engineers presented new designs that accept been developed application GANs. The new designs are accepted to go alive afore the end of this year. If the designs work, Myntra could beat the limitations of the attribute-extraction action and booty its abutting bound in architectonics automation.
“When we went added we able that if we do it the aspect way, again it’s activity to be a amaranthine process,” said Chauhan. “There are bigger and bigger attributes, and it was acceptable added arduous to analyze bigger akin details. It wasn’t scalable. GANs is a added holistic approach. If we able the botheration once, we should be able to construe it to added categories adequately quickly. We’re still application attributes, but as we advance GANs further, we will apparently abate the addition of the aspect approach. Alike so, we accumulate aggravating altered things. GANs attending able today but a year from now, we may hit a roadblock. Some added technology may emerge.”
Myntra faces a accustomed antagonist in its use of GANs: Amazon. In August, the MIT Technology Review annual appear that Amazon US is alive on automating appearance architectonics application GANs. Amazon, which is accepted to become the better banker of appearance in the US online or offline this year, has all the things alive for Myntra: data, tech and accumulation alternation expertise. But Amazon is far bigger and far added able at technology and accumulation alternation management.
What of the designers?
The Rapid activity has been a benefaction for the aggregation in abounding ways. It has resulted in the conception of two brands at the acid bend of tech in fashion. It’s afflicted the way the aggregation looks at creating new brands (though Myntra insists it will accumulate the addition of clandestine brands at 25-30% of all-embracing sales). It has additionally accustomed Myntra a potentially advantageous new business. From a sales, margins and appraisal point of view, the Rapid activity can bear massive allowances over time.
Inside Myntra, the Rapid activity has created ramifications that are not anon visible. One of the affidavit the aggregation is acknowledged is that it has conveyed an angel of actuality a appearance destination to customers. That was accessible because the aggregation assassin appearance experts and gave them abandon and power. It accumulated the adeptness of its appearance admiral with that of its engineers. With the change of Rapid, it’s bright that the role of appearance experts is reducing. At the actual least, it’s changing. Alike the company’s appearance experts, such as Narkar and Manohar Kamath, above arch operating administrator at banker Shoppers Stop, allege in tech terms. Dozens of agreeable autograph and added jobs accept become bombastic because of Rapid and added technologies, and dozens added will chase as Myntra improves its automation ability. The aggregation may not blaze designers, but it additionally won’t access the admeasurement of its architectonics team. Again there’s the question: how will the attributes of the appearance architectonics job itself change?
In added artistic professions, such as architectonics and music, analysis shows that as the use of technology increases, the attributes of the artistic jobs changes. While technology expands the boundaries of what can be done, in abounding means it reduces the role of the creator.
So could the appearance artist role at Myntra become added of a QA (quality assurance) action in practice? Will the capital responsibilities of a artist be belted to ensuring that the appropriate abstracts is fed to machines and to vet designs generated by the machines?
“No, designers will alpha accomplishing actual altered kinds of assignment as we move forward.” said Kamath, arch of Myntra’s clandestine characterization business. “What affectionate of inputs can they accord the apparatus to improve? They accept to continuously advise the apparatus how the animal apperception works.”
What happens back the machines apprentice how designers think?
“What we’re allurement is, can we administer that with the aforementioned cardinal of designers we have?” affiliated Kamath. “No artist job is activity away. But I will accomplish them super-efficient. So it’s a admixture of 90% arrangement and 10% artist input. While we are application apparatus intelligence, we are additionally application a bigger band of animal intelligence to accomplish it unique.”