AsLatticeFlowco-founderandCEOPetarTsankovtoldme,thecompanycurrentlyhasmorethan10customersinbothEuropeandtheU.S.,includinganumberoflargeenterpriseslikeSiemensandorganizationsliketheSwissFederalRailways,andiscurrentlyrunningpilotswithquiteafewmore.It’sthiscustomerdemandthatledLatticeFlowtoraiseatthispoint.
“IwasintheStatesandImetwithsomeinvestorsinPaloAlto,Tsankovexplained.“Theysawthebottleneckthatwehavewithonboardingcustomers.Weliterallyhadmachinelearningengineerssupportingcustomersandthat’snothowyoushouldrunthecompany.Andtheysaid:‘OK,take$12million,bringthesepeopleinandexpand.’Thatwasgreattimingforsurebecausewhenwetalkedtootherinvestors,wedidseethatthemarkethaschanged.”
AsTsankovandhisco-founderCTOPavolBieliknoted,mostenterprisestodayhaveahardtimebringingtheirmodelsintoproductionandthen,whentheydo,theyoftenrealizethattheydon’tperformaswellastheyexpected.ThepromiseofLatticeFlowisthatitcanauto-diagnosethedataandmodelstofindpotentialblindspots.Initsworkwithamajormedicalcompany,itstoolstoanalyzetheirdatasetsandmodelsquicklyfoundmorethanhalfadozencriticalblindspotsintheirstate-of-the-artproductionmodels,forexample.
Theteamnotedthatit’snotenoughtoonlylookatthetrainingdataandensurethatthereisadiversesetofimages—inthecaseofthevisionmodelsthatLatticeFlowspecializesin—butalsoexaminethemodels.
“
Henoted,forexample,thatmodelswilloftenfindhiddencorrelationsthatmayconfusethemodelandskewtheresults.Inworkingwithaninsurancecustomer,forexample,whousedanMLmodeltoautomaticallydetectdents,scratchesandotherdamageinimagesofcars,themodelwouldoftenlabelanimagewithafingerinitasascratch.Why?Becauseinthetrainingset,customerswouldoftentakeaclose-uppicturewithascratchandpointatitwiththeirfinger.Unsurprisingly,themodelwouldthencorrelate“finger”with“scratch,”evenwhentherewasnoscratchonthecar.Thoseareissues,theLatticeFlowteamsargues,thatgobeyondcreatingbetterlabelsandneedaservicethatcanlookatboththemodelandthetrainingdata.
LatticeFlowitself,itisworthnoting,isn’tinthetrainingbusiness.Theserviceworkswithpre-trainedmodels.Fornow,italsofocusesonofferingitsserviceasanon-premtool,thoughitmayofferafullymanagedserviceinthefuture,too,asitusesthenewfundingtohireaggressively,bothtobetterserviceitsexistingcustomersandtobuildoutitsproductportfolio.
“Thepainfultruthisthattoday,mostlarge-scaleAImodeldeploymentssimplyarenotfunctioningreliablyintherealworld,”saidSunirKapoor,operatingpartneratAtlanticBridge.“ThisislargelyduetotheabsenceoftoolsthathelpengineersefficientlyresolvecriticalAIdataandmodelerrors.But,thisisalsowhytheAtlanticBridgeteamsounambiguouslyreachedthedecisiontoinvestinLatticeFlow.Webelievethatthecompanyispoisedfortremendousgrowth,sinceitiscurrentlytheonlycompanythatauto-diagnosesandfixesAIdataandmodeldefectsatscale.”