Making Job Search Easier by Finding the Great Companies First

Find a
Title/Key­words Com­pa­ny Name
Where Search
City, state or zip (option­al)
City, state or zip (option­al)
Job title, key­words Com­pa­ny Name Only


Techcrunch Interviews: Infinite Analytics Can they Predict What you Will Buy Next?

If you want to work in the future today then this media job at Infi­nite Ana­lyt­ics could be for you.  Infi­nite Ana­lyt­ics enables their cus­tomers to use their eCom­merce data to pre­dict con­sumer behav­ior and per­son­al­ize expe­ri­ences. Infi­nite Ana­lyt­ics uti­lizes machine learn­ing, nat­ur­al lan­guage pro­cess­ing, ran­dom for­est mod­el­ing, image recog­ni­tion, and seman­tic tech­nolo­gies to dri­ve increased sales and con­ver­sions on client sites.

The secret sauce is match­ing indi­vid­u­al’s pro­files to prod­uct pro­files, which they call genomes.  In essence Infi­nite Ana­lyt­ics is learn­ing every­thing about the peo­ple that buy spe­cif­ic prod­ucts and match­ing that infor­ma­tion with the deep pro­files they devel­op of your web­site visitors.

We spoke with Arjun Mend­hi, the Direc­tor of Prod­ucts for Infi­nite Ana­lyt­ics, who talked about all the great MIT minds that are at work find­ing ways to max­i­mize sales from ecom­merce stores.

But isn’t all this per­son­al­iza­tion bias­ing the results?  Only show­ing your prod­ucts you’re most like­ly to buy vs let­ting you explore the prod­ucts?  Arjun explains how it all works to Roy Weiss­man of Media Jobs:



Roy: This is Roy Weiss­man from, and we’re talk­ing with Arjun Mend­hi, the Direc­tor of Prod­ucts for Infi­nite Analytics.


Arjun, tell us, what is Infi­nite Ana­lyt­ics? What do you guys do?


Arjun: Infi­nite Ana­lyt­ics is an ana­lyt­ics plat­form for eCom­merce indus­try. Essen­tial­ly, what that means is we enable prod­uct dis­cov­ery on the plat­form using var­i­ous tech­ni­cal advances that are com­mon­place today in var­i­ous oth­er indus­tries. Think of it this way. Today eCom­merce com­pa­nies spend a lot of time, effort and resources in attract­ing the right type of cus­tomers to their plat­form, but once they land at the plat­form, their expe­ri­ence, to say the least, it’s dif­fi­cult, because most of the eCom­merce plat­forms have way too many prod­ucts, and prod­uct dis­cov­ery is a chal­lenge. At every stage of nav­i­gat­ing through the web­site we observe that there are a sig­nif­i­cant num­ber of peo­ple that just drop off, so we address that prob­lem. We enable the cus­tomer to find the right prod­uct, and that’s our secret sauce, how we iden­ti­fy what would be the appro­pri­ate prod­uct for that par­tic­u­lar person.


Roy: Obvi­ous­ly, there’s a lot of tech­nol­o­gy out there, “Peo­ple who bought this also bought that,” or, “Peo­ple like this also like that.” What are you guys doing that’s dif­fer­ent, that’s not the sim­i­lar type of soft­ware that’s been out there for years?


Arjun: Sure. We’re not just dis­play­ing, “You like this. You’ll also like that,” just look­ing at click­stream infor­ma­tion. What we do is we focus on the machine learn­ing aspects to pro­vide a rel­e­vant expe­ri­ence all across the user inter­ac­tion. That means it could be some­thing as sim­ple as, “You may also like” to some­thing as com­plex as a per­son­al­ized search engine. This may take going out of our com­fort zone of real­ly envi­sion­ing what this looks like, but we’re so attuned to think­ing of a search engine as a key­word-dri­ven search based on a lot of pop­u­lar search engines that we use, that we don’t real­ize that there is a pos­si­bil­i­ty of a search engine being com­plete­ly per­son­al where two peo­ple who go to the same search engine, type in the same query, see com­plete­ly dif­fer­ent results, because they’re indi­vid­u­als with dif­fer­ent require­ments, espe­cial­ly on a shop portal.


How we achieve that is a two-part solu­tion. The first-part solu­tion is cre­at­ing a com­plete pro­file of every user, every vis­i­tor on an eCom­merce por­tal, whether it’s from his­toric click­stream infor­ma­tion to behav­ioral infor­ma­tion. We use all that data to cre­ate that pro­file. The sec­ond aspect, which is pret­ty unique to our solu­tion, is cre­at­ing a genome for every prod­uct on the cat­a­log, and that could be any­thing from mil­lions and mil­lions of prod­ucts to just a few prod­ucts, depend­ing on the cat­a­log size. We ingest such cat­a­logs and run them through our enrich­ment tools, and those enrich­ment tools have com­po­nents such as image analy­ses, where we iden­ti­fy pat­terns and col­or den­si­ties, to object recog­ni­tion, where they iden­ti­fy what this object actu­al­ly is, and cre­ate that meta­da­ta for the cat­a­log, which does­n’t exist.


You may send in a cat­a­log for a few giga­bytes, and we add meta­da­ta to it that enrich­es it and make it like a few ter­abytes pos­si­bly. Then that becomes the basis of this match­ing algo­rithm that we then run. Did that make sense?


Roy: Well, kind of. It sounds like it’s very per­son­al. There’s two things I want to ask.


Arjun: Sure.


Roy: One of them, which I’ll ask the sec­ond but I’ll get back to this in a sec­ond, just that, when you make some­thing very per­son­al, isn’t it almost bias­ing the results to a self-ful­fill­ing proph­esy? Don’t answer that until you answer this. You men­tion machine learn­ing. We’ve heard a lot about machine learn­ing. We’ve heard about arti­fi­cial intel­li­gence. How would you define machine learn­ing, and what is the dif­fer­ence between machine learn­ing and arti­fi­cial intelligence?


Arjun: Sure. In very sim­ple terms, arti­fi­cial intel­li­gence is the user infrac­tion lay­er, and machine learn­ing is the back­end sys­tem that makes arti­fi­cial intel­li­gence pos­si­ble. When you inter­act with a sys­tem that seems to be intel­li­gent, it’s called the Tur­ing test, where you inter­act with the sys­tem, and if you’re unable to dis­tin­guish or clear­ly iden­ti­fy whether you’re inter­act­ing with a per­son at the oth­er end or a machine, then that is arti­fi­cial­ly intel­li­gent as a sys­tem. That inter­ac­tion lay­er, I believe, is arti­fi­cial intel­li­gence, and the sys­tems that per­form the learn­ing and train­ing of mod­els is the machine learn­ing layer.


The sec­ond part of your ques­tion, could you repeat that?


Roy: With the machine learn­ing, you’re say­ing the machine learn­ing is a pri­ma­ry com­po­nent of arti­fi­cial intelligence.


Arjun: Yes.


Roy: It’s a sub­set, but arti­fi­cial intel­li­gence is a big­ger pic­ture. It has more capa­bil­i­ties. Is that what you’re say­ing, the machine learning?


Arjun: No. I’m say­ing there are two lay­ers to an appli­ca­tion, an intel­li­gent appli­ca­tion. One is the lay­er that learns from train­ing data or learns, that learn­ing can be super­vised or unsu­per­vised. That learn­ing lay­er is the machine learn­ing lay­er. Once you have mod­els that have that learn­ing and the train­ing embed­ded in them, you can cre­ate appli­ca­tions on top of them which are seem­ing­ly intel­li­gent. That is arti­fi­cial intelligence.


Roy: Okay. Machine learn­ing is just learn­ing it, and intel­li­gence is apply­ing that learning?


Arjun: Yes, you can say that.


Roy: Okay. The oth­er ques­tion I ask is we always talk about per­son­al­iza­tion, and Google has tried to do some of that by when you sign into your account it keeps track of the kinds of things you like to see, so it tries to show you search results that relate to things you want to see. Some­times when you talk about per­son­al­iz­ing search based on my behav­iors on a web­site or what­ev­er, one of the things that intrigues me about this is I always feel like aren’t you effec­tive­ly cre­at­ing a self-ful­fill­ing proph­esy? You’re lim­it­ing the results to what the com­put­er thinks you want to see.


It’s kind of like hav­ing an office full of “Yes” men. Nobody every dis­agrees with you, so you start to think you’re right. If you start search­ing for some­thing and it thinks, “Oh, well, he real­ly wants golf shoes,” and he starts show­ing me golf socks and golf shoes, and I’m search­ing for some­thing dif­fer­ent. I just hap­pen to ini­tial­ly out of curios­i­ty look because a friend of mine bought golf shoes. Aren’t I lim­it­ing my results? Isn’t this per­son­al­iza­tion lim­it my abil­i­ty to max­i­mize, or do you find a dif­fer­ent val­ue to it?


Arjun: Yes, that’s a great point. That’s also the bound­ary between being sup­port­ive as plat­form, being help­ful, and being creepy. That’s some­thing we think of very care­ful­ly in how we design our sys­tems. The best way I can put it is we make the process of dis­cov­er­ing the right prod­uct a seam­less expe­ri­ence and not an intrud­ing expe­ri­ence. What that means is we take live cues from what the user is doing and sup­port that process as opposed to make our assump­tions and con­scious­ly direct the con­sumer in a direc­tion that the con­sumer may not want to go into. We’re not walk­ing in with a pre­dis­po­si­tion. We’re not try­ing to dri­ve the con­sumer in a direc­tion which is dis-joined from the con­sumer’s inten­tions. This brings me back to our machine learn­ing sys­tems, where we’re able to match the user’s real-time behav­ior with what we think would be a com­ple­men­tary show­case of prod­ucts, and that’s def­i­nite­ly based on data.


To sum­ma­rize, yes, that is a con­cern, but I think it’s a del­i­cate bal­ance between whether you are impos­ing prod­ucts that you actu­al­ly want the user to look at or you’re just let­ting the data do its job and you’re pro­vid­ing the user insight into prod­ucts that might have prob­a­bly tak­en him or her var­i­ous clicks and var­i­ous oth­er pages and search­es to get to.


Roy: No, that makes sense. I think that makes a lot of sense. When we talked ear­li­er, you had men­tioned that you guys had done some tests, and had been a com­pe­ti­tion or some­thing where they put your soft­ware up against oth­er soft­ware. Maybe you could just explain that a lit­tle, the results.


Arjun: Sure. That’s a sim­ple AB test where with­out giv­ing specifics of the client or the oth­er con­tenders, it’s essen­tial­ly one of the biggest eCom­merce plat­forms in the world. The plat­form had con­scious­ly decid­ed to improve its per­son­al­iza­tion aspects and brought in some of the lead­ing ven­dors in the world, includ­ing the incum­bent play­er. We had ABC essen­tial­ly, an ABC test, with the three short-list­ed ven­dors get­ting a por­tion of that traf­fic. The test ran for three weeks live on their web­site, and every aspect of the per­for­mance of these per­son­al­iza­tion sys­tems was mon­i­tored. We were by far the small­est com­pa­ny in the com­pe­ti­tion. The oth­ers were fair­ly well-estab­lished, but we won that com­pe­ti­tion, that test, in every sin­gle para­me­ter that had a busi­ness sig­nif­i­cance to the por­tal that includ­ed met­rics such as aver­age order val­ue, total rev­enue gen­er­at­ed from our rec­om­men­da­tions and per­son­al­iza­tion engines, total traf­fic gen­er­at­ed, click-through rates and things like that.


Again, it’s impor­tant to remem­ber that these sys­tems are, they’re basi­cal­ly machine learn­ing sys­tems which get bet­ter and bet­ter with more and more train­ing data. If you bring in an out­side enti­ty and give them a small per­cent­age of your data and put them live on your web­site and com­pare that engine with the incum­bent engine, which has been on the web­site for, say, years, then there’s a huge dif­fer­ence in how well-cal­i­brat­ed or learned these mod­els are. Despite that dif­fer­ence, we were able to win the competition.


Also, one of the biggest chal­lenges we had was that the por­tal, the eCom­merce engine, had cus­tomers that were using the por­tal in a non-Eng­lish lan­guage, so not only did we have to under­stand the behav­iors of a com­plete­ly dif­fer­ent demo­graph­ic than our core mar­kets, we also had to trans­late and under­stand meta­da­ta in a dif­fer­ent lan­guage. Despite all these chal­lenges, we were very excit­ed that we won that test, and it speaks vol­umes to our unique approach and how that is not just the­o­ret­i­cal­ly supe­ri­or but also deliv­ers results.


Roy: Did you have any cal­cu­la­tions of how much that could have been worth to the com­pa­ny on a glob­al basis or in dol­lars and cents or some­thing like that?


Arjun: Yes. The com­pa­ny them­selves did that. They were mon­i­tor­ing it very close­ly, and they were extrap­o­lat­ing the frac­tion­al data that we were get­ting to a sce­nario where each com­pa­ny would get 100% of the data. Giv­en that, yes, all I can say is, it was worth a lot. It was worth a sig­nif­i­cant chunk to the orga­ni­za­tion. They’re still work­ing on final­iz­ing their next steps, so it’s a fair­ly recent test that we came out of.


Roy: That’s very exciting.


Arjun: Thank you.


Roy: Hope­ful­ly, you’ll be able to share that with a few oth­er poten­tial cus­tomers, those kinds of results.


Arjun: We’re hop­ing to, yes. Once this is pub­lic, we are hop­ing to also share more infor­ma­tion about it. Check back with me in a few weeks.


Roy: Do you guys have any sense of the size of the poten­tial … Obvi­ous­ly, you’re going to license your soft­ware, I assume, and peo­ple will pay you a fee. Is that going to be based on a per­cent­age of rev­enue, or it’s just a fixed fee? How do you envi­sion pric­ing it?


Arjun: Yeah, that’s a great ques­tion. One of the things we observed about this mar­ket is that pric­ing is very opaque, and it’s very, very cus­tomized and extreme­ly nego­ti­at­ed. One ven­dor has no clue how the mar­ket is pric­ing such ser­vices, so we try to take a com­plete­ly dif­fer­ent approach. Our approach is as trans­par­ent as it gets. We talk about a val­ue we can deliv­er. We ana­lyze their exist­ing per­for­mance, and we look at what’s that delta that we can create.


Let’s say cur­rent­ly take a hypo­thet­i­cal sce­nario where a cus­tomer has a rec­om­men­da­tion engine that’s pro­vid­ing, say, X per­cent­age of their rev­enues through cer­tain click-through rates to cer­tain aver­age auto val­ue. We com­pare that with their gener­ic traf­fic, and so if the oth­er data that we have from our oth­er engage­ments, we know with very lit­tle or, let me say, with a well-defined mar­gin of error, we know what is the dif­fer­ence in per­for­mance we can deliv­er. We take that dif­fer­ence in per­for­mance that we can cre­ate in the top line for the cus­tomer as the total val­ue cre­at­ed, and we cap­ture a cer­tain por­tion of that val­ue for our ser­vices. The rest of it is left on the table for the cus­tomer, which is a com­fort­able margin.


Now we give the cus­tomer a lot of flex­i­bil­i­ty in how they want to struc­ture that val­ue that we cap­ture. They can struc­ture it in var­i­ous ways. They can award that to us as a flat month­ly fee, which of course is pret­ty trans­par­ent. If X dol­lars is being cap­tured by our com­pa­ny, that can be divid­ed by the dura­tion of the engage­ment. Should they choose to make it per­for­mance-based, a lit­tle more per­for­mance-based, they can split that between a small­er por­tion as a month­ly fee and a cer­tain per­cent­age as a rev­enue share, as a rev­enue share of the total rev­enue gen­er­at­ed by the rec­om­men­da­tion engine, or if they want to make this a hun­dred per­cent per­for­mance-based, our entire com­pen­sa­tion can also be as a rev­enue share based on the rev­enue we gen­er­ate for the customer.


There are var­i­ous levers that we offer our cus­tomers to work with. The idea is essen­tial­ly pret­ty sim­ple. We know what’s the val­ue we can cre­ate, and we want to hold our­selves by those results. We under­stand every busi­ness is dif­fer­ent, and their way of work­ing with sup­pli­ers and ven­dors is dif­fer­ent. We want to be as flex­i­ble as pos­si­ble while being fair on both sides. We don’t see any rea­son not to be trans­par­ent in our pricing.


Roy: Have you got a sense, have you guys sat down and esti­mat­ed the size of the mar­ket oppor­tu­ni­ty for you guys in dollars?


Arjun: We con­stant­ly try to do that, but being a start­up, we are also very adapt­able. At some point we can make big state­ments about how such per­son­al­iza­tion indus­try is a mul­ti-bil­lion-dol­lar indus­try with more than 30 ven­dors try­ing to claim the space, includ­ing larg­er orga­ni­za­tions like IBM and small­er orga­ni­za­tions like, but I think we try not to get swayed by such num­bers. We try to be very, very focused on what our dif­fer­en­ti­a­tion is and who these cus­tomers are that will val­ue our ser­vice. The short answer is, yes, there are var­i­ous pub­lic reports which give large num­bers, but I don’t think we actu­al­ly get influ­enced by those. We take one cus­tomer at a time.


Roy: What do you think the poten­tial could be, dollars-wise?


Arjun: We are pret­ty sure the path that we are on right now would help us become a bil­lion-dol­lar com­pa­ny in a not so long future.


Roy: Wow.


Arjun: Let me put it that way. It’s not just the, as we talked about, it’s not just rec­om­men­da­tion engines. It’s a lot more than that. It’s the intel­lec­tu­al prop­er­ty and the IP that we are cre­at­ing in the com­pa­ny which has a lot of var­i­ous appli­ca­tions that we are not even explor­ing at the moment. We’re try­ing to put one step ahead of the oth­er in a very strate­gic man­ner, excel at what we do first before we start diver­si­fy­ing, but def­i­nite­ly diverse appli­ca­tions are down the roadmap.


You can think of var­i­ous ways machine learn­ing can ben­e­fit dif­fer­ent busi­ness­es in the world, whether it’s any­thing from drug dis­cov­ery appli­ca­tions to health­care to finan­cial ser­vices. There are var­i­ous, var­i­ous appli­ca­tions to the tech­nolo­gies that we’re cre­at­ing. Right now our focus is, deliv­er val­ue in the indus­try we play in to the cus­tomers we work with and be the world’s best at that. Once we’ve cre­at­ed that intel­lec­tu­al prop­er­ty, we can cre­ate the room to explore diverse appli­ca­tions, but right now we’re extreme­ly focused.


Roy: When were you guys founded?


Arjun: Two thou­sand twelve, out of MIT.


Roy: Had the prod­uct offi­cial­ly been released yet, or is it still being test­ed at this point?


Arjun: Oh, the pro­duc­t’s been released. It’s live on var­i­ous sites with var­i­ous cus­tomers. One of our strate­gic deci­sions ear­ly on in the com­pa­ny was to start with emerg­ing economies. First of all, they are see­ing just a boom­ing growth in eCom­merce. It’s also rel­a­tive­ly eas­i­er to exper­i­ment in emerg­ing economies with com­pa­nies that are more open to nascent and just tech­nolo­gies that are still under development.


We saw a lot of trac­tion in emerg­ing economies. Cur­rent­ly we have some of the largest plat­forms, eCom­merce plat­forms, say, from India, work­ing with us. We have pen­e­tra­tion in Brazil. We believe we’ve reached a point where we’ve proven our prod­ucts and our tech­nolo­gies to be the most supe­ri­or in the world in var­i­ous aspects, and we feel we’re ready to enter the US mar­ket now. Now is the time where you start see­ing the name “Infi­nite Ana­lyt­ics” be more preva­lent in the US mar­ket and media. We’ve tak­en the time to not rush into this mar­ket with a half-baked solu­tion. We think we’re ready with a solu­tion that can com­pete with the best in the world, and we’re optimistic.


Roy: Who found­ed the company?


Arjun: Two grad­u­ates from MIT, one of them in a tech­ni­cal grad­u­ate course, and one of them in the MBA pro­gram com­ing out of the school in 2012. It’s an inter­est­ing sto­ry how they actu­al­ly came up with the first ver­sion of this product.


Roy: Could you tell us the names?


Arjun: Yeah. Akash Bha­tia, the CEO.


Roy: Okay.


Arjun: Purushotham Bot­la, the CTO in the com­pa­ny. They met at a class taught by Sir Tim Bern­ers Lee called Linked Data Ven­tures, and Linked Data Ven­tures essen­tial­ly talks about the future of the inter­net, the future of the world­wide web, the seman­tic web and linked open data, and the val­ue that can be cre­at­ed from sys­tems that uti­lize this future vision of the web. They start­ed out with a class project called Schmooze-But­ler, which was essen­tial­ly a net­work­ing appli­ca­tion on social media, but fun­da­men­tal­ly it con­nect­ed dis­parate data sys­tems across the web in a very intu­itive man­ner that opened a lot of room for ana­lyt­ics and appli­ca­tions. That was so impres­sive that they decid­ed to launch a com­pa­ny on the basis of that project, and till today Tim Bern­ers Lee is an advi­sor for Infi­nite Ana­lyt­ics, so that speaks to the trac­tion we’ve cre­at­ed and the vision we have.


Yes, that’s the foun­da­tion. We’re still look­ing into the future. We always look at where inter­net is going to be as a whole ten years down the road and not one or two years down the road. That’s always been a dri­ver for our prod­uct strategy.


Roy: How many peo­ple do you have work­ing there now?


Arjun: I’m look­ing around our office right now. We have about 15 peo­ple here in Cam­bridge. A lot of interns have just joined from MIT. More to come next week. We have about five peo­ple in Europe. We have about five peo­ple in India right now.


Roy: You’ve got about 25 now.


Arjun: Yes.


Roy: Are you look­ing to hire people?


Arjun: Yes. We are cer­tain­ly hir­ing. We’re hir­ing for dif­fer­ent roles. On the tech­ni­cal side, we’re look­ing for sys­tem oper­a­tions man­agers who can help us scale our infra­struc­ture. We’re look­ing for per­for­mance engi­neers. Our sys­tems are com­plete­ly real-time with very, very ambi­tious response times of sub-20-mil­lisec­onds for var­i­ous appli­ca­tions, so per­for­mance engi­neer­ing is a very impor­tant role in the com­pa­ny. We’re look­ing for data sci­en­tists. We have a very high den­si­ty of data sci­en­tists in the com­pa­ny already. That’s essen­tial­ly the foun­da­tion of every­thing we do. Then we’re look­ing for UX devel­op­ers, user expe­ri­ence engi­neers, who can help us build cus­tomer-fac­ing prod­ucts that are seam­less, intu­itive and best in class in the world. Then on the busi­ness and man­age­ment side we’re look­ing for prod­uct man­agers, and we’re look­ing for busi­ness devel­op­ment asso­ciates to join our team. In busi­ness devel­op­ment we’re look­ing for peo­ple both in the US and in Brazil.


Roy: It sounds like you’ve got a lot of things going on. It’s about to explode, huh?


Arjun: Yes, it’s an excit­ing time, so stay in touch.


Roy: If I come back in five years, will you still be an inde­pen­dent com­pa­ny? Will you be a pub­lic com­pa­ny, or will you have been acquired by some big company?


Arjun: Of course we don’t know the future, but what I can tell you pret­ty con­fi­dent­ly is an exit has nev­er been a dri­ver in the com­pa­ny. We nev­er do any­thing to make us more attrac­tive for a cer­tain exit oppor­tu­ni­ty. We nev­er think of acqui­si­tions. We nev­er think of just exit­ing in a fash­ion. We think of just excelling at what we do. We think of how we can build our tech­nolo­gies and our team to cre­ate max­i­mum growth in the com­pa­ny. I think being dis­tract­ed by such dri­vers could be detri­men­tal at an ear­ly stage in the company.


Roy: Is there any­thing I haven’t asked you that you want­ed to share with the audience?


Arjun: Just that we’re based out of Cam­bridge in Mass­a­chu­setts. That’s our head­quar­ters. If any of your audi­ence are in Cam­bridge or Boston and would like to touch base with us, we’re a great place to work. We work in a shared open envi­ron­ment with a lot of oth­er star­tups. It’s a lot of fun work­ing here. There are pret­ty fre­quent ping-pong tour­na­ments. There are team hang­out lunch­es. Cam­bridge is a very intel­lec­tu­al and aca­d­e­m­ic envi­ron­ment with a lot of stu­dents, a very vibrant com­mu­ni­ty. It’s a great place to work.


If any of your audi­ence mem­bers that’s around wants to meet us, then just shoot us a note at or go to our web­site and shoot us a note. We’re always hap­py to con­nect with peo­ple who are eager to con­tribute or just learn more about us.


Roy: The web­site is


Print Friendly, PDF & Email