Who couldn’t use some time-lapsing, truth-seeing artificial intelligence models at work? Such models analyze whole object-storage buckets full of text or images. They answer difficult queries in minutes, sparing users weeks or months of tedious work. Sign me up.

The trouble is, not all companies have a dream team of data-science Ph.D.s on hand to tailor custom AI solutions. Vendors are increasingly promising instant insight via applications with baked-in AI. But the latter might not be as spot-on as models custom-built by in-house experts.

Is there any middle ground offering customization with a barrier to entry that data laymen can clear? There is — if users can manage just a smidgen of coding.

DimensionalMechanics Inc. allows customers to build highly sophisticated AI models on its platform. It has its own built-in AI, dubbed “the oracle” — DimensionalMechanics is filled with fervid fans of The Matrix. The oracle guides users as they build their models.

“It has a vast knowledge base,” said Rajeev Dutt (pictured), co-founder, president and chief executive officer of DimensionalMechanics. “It has a lot of additional machine-learning components and things like that that essentially allow it to adapt and learn based on the kind of problem you’re trying to solve.” 

 The word “language” can instantly scare off people with zero machine-learning or coding experience. But DimensionalMechanics’s NeoPulse Modelling Language, or NML, is nothing they’ll need to go back to college for, according to Dutt.

He has seen people with no tech chops, like university professors, build AI models with NML. Sometimes they get the job done with as few as 14 lines of code. “We had a high school student who spent about a week learning it,” he said. “A week later, she was ready to start coding, and she had built her first models using that.”

Dutt spoke with Jeff Frick, host of theCUBE, SiliconANGLE Media’s mobile livestreaming studio, during the AWS Marketplace and Service Catalog Experience Hub event in Las Vegas. They discussed how businesses are using DimensionalMechanics’s platform to build accurate AI models more cheaply. (* Disclosure below.)

This week, theCUBE spotlights DimensionalMechanics in its Startup of the Week feature.

Making moves to AI endgame

The big data trend that exploded in the industry several years ago left many disappointed. The software that vendors and open-source communities brought out didn’t yield jaw-dropping gains for the majority of users. Today, 85 percent of big data projects in enterprises fail, according to Gartner analyst Nick Heudecker‏.

So why do enterprises keep hacking away at such projects? Because the potential of big data to improve all kinds of business outcomes is clear. Now the industry at least has a better picture of what successful big data projects look like. It’s not just a data lake; it’s not a collection of halfway-decent analytics tools. The endgame of big data is AI — real, instant answers to questions via trained models and appropriate actions triggered by data, information and events.

We’re now seeing technologies bring users closer to the AI endgame right out of the gate.

“One of the most disappointing things to me in our industry is that most of the AI projects have boiled down to a [crappy] chat bot,” Matt Lancaster, tech architecture science associate director at Accenture LLP, recently told theCUBE. “They can’t actually do anything.”

Event-driven architectures that react to single events rather than complex sequences can deliver AI for better end-user experiences, according to Lancaster.

Solutions that allow users to attach AI to applications like Lego blocks are also surfacing. Amazon Web Services Inc. made a number of announcements around easy machine learning and AI add-ons at re:Invent in Las Vegas in November. (DimensionalMechanics is an AWS partner and comes up in the AWS Marketplace’s searchable ML subset.)

“You now have a simple [application program interface] to basically build these modern AI-powered apps,” Jerry Chen, partner at Greylock Partners, recently told theCUBE. “In the future, can you build a [software as a service] application entirely on Amazon, Azure or Google without any custom code?”

Perhaps. For now, some DimensionalMechanics users are building highly accurate AI models for business use cases. The way they hug the curves of their unique problem and data make a dash of coding easily worth the effort.

Pro-strength AI without the CS degree

NML can chop off 85 percent of code length compared to deep-learning architectures written in Python with Keras, according to Dutt. “My wife is a radiologist, and she’s actually looking at using it for her own internal research projects,” he said. 

Training from scratch allows a high degree of customization. DimensionalMechanics also has features useful for radiology models that are not covered in standard image-recognition AI. “Using things like transfer learning or fine tuning doesn’t help in this particular case, because if you’ve trained a model in dogs and cats, then training it to recognize stress patterns in a hull is just not going to work,” Dutt explained.

The AI on DimensionalMechanics’s platform helps lay people make appropriate choices about their models. It might make suggestions about functions, hyper parameters, optimizers and the like.

Users can export the portable inference model object and deploy it to any target as long it’s on DimensionalMechanics’s run time. This could be cloud, a field-programmable gate array at the edge, etc.

DimensionalMechanics users have seen huge leaps in accuracy when switching to its platform, according to Dutt. Models usually give a pretty accurate answer the first time users try them, he added. They learn more and become more accurate as users retry them again and again.

One customer, a resume-matching company, had a prospective vendor quote $450,000. “They were warned that they would not be able to exceed 40-percent accuracy given the data that they had,” Dutt said.

With DimensionalMechanics, it achieved 83 to 84 percent accuracy for under $10,000. Another customer was spending $20,000 per month on a technology to measure heart rate. DimensionalMechanics brought it down to $4,000. 

Stanford University physicians with no coding or engineering skills built a model for PET-CT, a 3-D image of the human body. It determines whether someone has a tumor or not. With a limited data set, they achieved 74 to 75 percent accuracy. The model was good enough to publish at RSNA, one of the biggest radiology conferences in the world, Dutt said.

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of the AWS Marketplace and Service Catalog Experience Hub event. (* Disclosure: TheCUBE is a paid media partner for the AWS Marketplace and Service Catalog Experience Hub event. Neither Amazon Web Services Inc., the event sponsor, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE