The company now has received $90.3 million in financing. Tiger Global is joined by returning investors Coatue Management and Index Ventures, as well as Alkeon. 
“A large chunk of it will goes to R&D and engineering and science,” explained Bindu Reddy, co-founder and CEO of the company, in an interview with ZDNet via Zoom. “We continue to want to be the best of breed in AI and ML platforms.” The other main use of the capital will be for go-to-market, including building out the sales team. See also: AI startup Abacus.ai turns on real-time deep learning systems for enterprises. Abacus currently has 45 people, including four people in sales and marketing. Reddy expects to expand the total company headcount to 80 by the end of this year. 
Along with the financing news, Abacus.ai unveiled a computer vision-as-a-service application. The company had begun with applications for tabular data, following that earlier this year with natural language processing applications. 
“The main difference between us and anyone else in the market today is hybrid models,” said Reddy. “If you look at someone like OpenAI or even Google, they have language models, they have vision models, but these are pure language or vision models,” in which the function, such as image classification, is narrowly tuned to the particular data type. 
“What we are doing is support for hybrid models where you can combine language, vision, and structured data to get better results on your models.” An example, said Reddy, would be a finding the price of a home, not just based on feature attributes, such as the number of bedrooms, but also the description in natural language, and then the photographic data that shows qualities of the home. See also: AI startup Abacus.ai nabs $22 million in Series B funding to automate the creation of deep learning models. “Adding that language and vision signal into a predictive model is what we are focusing on, and what we end up being very good at.”
The customer example sounds reminiscent of what Opendoor does with deep learning, for example. Reddy confirmed the home sales application is a real customer application but declined to identify the customer. 
To date, over 10,000 customers have used Abacus.AI to train over 30,000 models, the company says, “and several of them, including 1-800-Flowers, Flex, Recorded Books, Daily Look, and Prodege use Abacus.AI in production for several of their AI use-cases.”
The appeal of hybrid models is two-fold, said Reddy: it is more resource-efficient than the classic deep learning approaches involving very large numbers of parameters, and most enterprise problems really resemble hybrid approaches, she said. 
“We are a startup, so to do pure vision or pure language models from the ground up takes a lot more money,” observed Reddy. 
“And honestly, most enterprise use cases are hybrid – we are more or less doing applied AI.” Feature sets are more often curated by an enterprise user rather than massive amounts of feature discovery.  See also: AI startup Abacus goes live with commercial deep learning service, takes $13M Series A financing. “Our whole goal, at a very high level, is to say, have we extracted all the intelligence from this data.”
Abacus.ai is cheaper than cloud vendors for AI applications, argues Reddy, because of the focused nature of the enterprise use cases. “The issue with most cloud platforms is that you spend a lot of money experimenting, more so than putting things into production,” she said. “The experimentation needed to put things into production is much less” with the Abacus.ai applications, she said.
Abacus.ai uses a consumption-based pricing scheme, where one pays by the number of predictions being made. That is similar to Snowflake in the data warehousing market. “The problem with licensing is that they tend to be not value-driven, and it stops adoption in some ways, by telling people they have to spend more money to adopt the product.” 
In the same way that Snowflake optimizes on data compression, said Abacus.ai, Abacus can reduce costs for customers. “We do a whole bunch of optimization of the models that we’re running so that we don’t have to spend too much for compute.”  “You can pack models together to be on the same server, you can run a hotspot cluster for data transformations,” are examples of cost savings. In addition to running a commercial enterprise, Abacus.ai researchers continue to publish in academic circles on deep learning computer science. Abacus.ai has various research papers that have been accepted for conference publication. Two are about the company’s core competence in “neural architecture search” for discovering the optimal architecture for a neural net automatically. Those papers have been accepted at this year’s NeurIPS AI conference.      Other papers pertain to explainability, said Reddy, including one accepted at NeurIPS regarding benchmarks for explainabilty; and another that addresses explicitly tabular data, called “Regularization is all you need.” 
Another paper under development will focus on the hybrid approach.