CNBC.com states, "Tara AI faces competition from large tech consultancies like Tata or Insight Global, and freelancer and developer marketplaces online like Upwork, Codementor and Gigster." For this reason, how Tara AI works can only be conjecture, as they would be doing all they can to plug any leaks of company confidential information. There won't be any papers coming from their AI experts that describe their matching algorithms and data collection and normalization techniques.
Matching a worker to a project is a matter of scanning a large number of documents about workers with a requisition from the recruiter or recruiters assigned to staff the project. The approaches to apply AI to this problem that could produce favorable results are numerous. It is likely that many of these organizations listed by CNBC.com like to say to experts, recruiters, and project management that they use AI in the matching, even if they are merely using a probabilistic model based on three things.
- Frequency of words on the web (ambient word frequency)
- Frequency of words in the resumes
- Frequency of words in the requisition
Beyond that high level description, I too am constrained by information privacy agreements.
Certainly, semantic nets is another approach that may be in use in some of these organizations, although there is much to do in the field of AI, with the mathematics, the hardware, and the software, before we have cognition components to use that actually understand what, "Full Stack Java Developer," means to the degree that the expert and the team leader who writes the requisition understand it.
It's a set of tricks people use, like the tricks that are used in 3D rendering. The pixels don't have to be perfect ray traced from every light source and reflection for every frame in the render. Just close enough so that viewers are not distracted by digital artifacts is sufficient. With matching people with roles, there is no way to gather enough data to perform an highly reliable prediction.
Due to the complexities of a billion brain cells times however many people there are on the project team, some people just fit into the role in spite of predictors otherwise, and vice versa. That is simple predictive tricks may produce a reliability of placement very close to the use of a deep learner running two floors of data center.