What’s your background?
I was fifteen years with Microsoft working on semantic search, knowledge management and reasoning, on Satori and Cortana. At Maana we are now working on the application of semantics in the enterprise, using it to ‘understand’ data, both big and small.
When you say semantics, do you mean the semantic web?
No the semantic web is essentially a failed technology, an academic effort which had some limited success in pharma and healthcare. It’s capacity for machine readability and inference was promising but the effort required to prep data for such applications was unsurmountable. We tried in at Microsoft but it is not practicable for an unknown, ad hoc domain. At Maana, semantics means a blend of data structure and statistics. In healthcare, you might know that a certain drug treats a specific disorder. But what is of real interest is how often it is taken, who’s taking it and what the outcomes are. Here, machine learning is a great way of leveraging both the statistics along with the structured data.
You use the Accumulo graph database.
The graph database is key to our approach but we have moved on from Accumulo. We now work with a proprietary store that embeds our patented graph-based ‘dynamic semantic model.’ We still use the Hadoop file system and we support Spark. Our system provides a fluid data representation that builds on category theory, the math that underpins our technology.
So this is a shift from open source to proprietary software?
Yes, for storage and for scale and efficiency. We still have extensive support for Spark and R, in fact we contribute to these projects and we also work with Hortonworks and MapR.
So how does it work?
Our flexible representation of data allows us to remap information into ‘kinds.’ Thus we repurpose and link data say from well to geoscience or from well to finance, changing the perspective of the model in a computationally efficient manner. Maana also allows ad-hoc filtering and data structures and indices that are tuned to different data types. This includes text with substring matching and time based data. This might connect a kick event in real-time data with text recorded within a certain time frame, rolling-in Bayesian inference.
‘Kinds’ sounds like business objects?
Sure, or ‘views.’ For us a ‘kind’ is our version of concept variation.
Accessing different data sources sometimes comes unstuck on issues like well names in difference source databases. How do you handle this?
There is no magic involved! You need to align terms in different databases. Our machine assistance can help here, guiding and learning from users as they map field names across different systems.
Can you give some use cases?
The front end takes raw data from various APIs into line of business apps for say, field services, such as part ordering by field teams to minimize returns. Maana’s enterprise architecture is integrated with the business. You can type in device/error codes or symptoms and the system advises on remediation. We call this ‘data driven device decision support.’ Another tool provides predictions on accounts receivable, predicting when an invoice will be paid. Others ingest well data sets, match on the well name and see what happened following stimulation, rotating views and traversing the graph.
GE (an investor in Maana) talks about smart preventative maintenance. Has Maana displaced GE’s internal tools?
No. But we are working with Predix and with GE Digital on this kind of thing.
What about upstream data sources? Companies have spent years building connectors for hundreds of data sources.
Database toolbox connectors and OBDC does fine for most of our stuff. We outsource connector development on an as-needed basis.
Do you work downstream of the historian?
Certainly, some clients have large elaborate data lakes. We support the business ecosystem, Tableau, Spotfire… We elevate these systems to the level of a canonical source of knowledge.
What does Maana mean?
It is Urdu for ‘meaning.’
More from Maana.
© Oil IT Journal - all rights reserved.