Deep Learning Development
Client is a medical technology company making tools for processing heartbeat data.
Although they did have data-scientists on the team, their team was not familiar with deep-learning and needed an expert consultant for a variety of ad-hoc tasks. Luckily they found Electric Brain to help them out. Initially, Electric Brain just met weekly with the Analytics4Life team and provided quick consultations to assist them with their internal efforts.
Later, Analytics4Life needed a script to test out various deep-learning architectures on their data. They had a tight deadline and Electric Brain was able to step in and provide assistance creating the script on short notice so that they could hit their deadline.
However, they couldn’t get enough and eventually requested Electric Brain assist them with everything from deep-learning visualization to testing out capsule networks.
Electric Brain was able to provide a quick and timely addition to the skills of their team, allowing them to explore deep-learning faster and hit their tight timelines.
Blueprint Software Systems is a company located in Toronto which creates a program called Storyteller, used for designing and managing IT projects. The software is sold to large enterprises, who use it to manage their increasingly complex IT systems and projects in an agile manner.
One of the big pain points for Blueprint’s customers is compliance. The Storyteller makes this easier by encoding many different kinds of regulations into small, terse sentences called “controls”. It then allows users to tag parts of their project plans with relevant controls.
For example, if a part of a project plan suggests “A user should be able to download their own transaction data, so that they can import it into a personal finance application”,
then you might tag this with a control that says “All data
must be encrypted when transmitted across the internet.” This process is tedious and painful, and Blueprint’s talented product manager, Tony Higgins, believed that automating this process would offer tremendous value to their customers, saving up to 90% of the time required to do compliance review.
Blueprint first experimented with developing the technology internally with their existing IT team, testing techniques such as keyword matching and logistic regression analysis. However, these techniques proved to be insufficient and Blueprint’s team lacked the experience in machine learning to develop more advanced techniques. This is where Blueprint hired the AI consultants at Electric Brain to help research a solution.
Electric Brain came up with a 4 step plan to build out a solution:
Build up a dataset of user-stories and their corresponding regulations
Researching machine learning algorithms using that dataset
Integration of the algorithm into their product
Monitoring and tuning of the performance after launch
Blueprint decided to move forward with a proof-of-concept, taking the first two steps towards building their AI system. Blueprint rallied its clients and obtained a small dataset that could be used for testing. Electric Brain then got to work researching deep-learning technology on this dataset.
After much experimentation, taking advantage of the latest machine learning techniques, Electric Brain discovered a 4 layered recurrent neural network that performed effectively on this task.
It was proven, for the first time, that it’s possible to automatically detect relevant regulations and controls for parts of IT project plans. This validated the concept and was a tremendous first step towards automating compliance review using AI technology.
The project is now proceeding to the second phase. Blueprint is procuring a much larger dataset from its clients, from which we will train a joint deep neural network and integrate it into the product. Implementation of the technology is expected to save Blueprint clients a total of $1.3 million dollars per year in their efforts to comply with government regulation.
HR Proof of Concept
The client is an entrepreneur and owner of a recruiting firm. He is interested in starting a new business to offer AI powered recruiting.
The first step in building the business was to show the viability of applying AI tech to just a single part of the recruitment process: screening. The client would build this as their MVP and ultimately launch it into the marketplace. So to test whether AI technology was viable, the client approached Electric Brain to build the proof of concept.
Electric Brain worked with the client to extract a dataset of resumes from their prior recruiting efforts.
Electric Brain then went to work testing out different algorithms that could be used for AI based resume screening. After a few hiccups testing out different kinds of deep-learning architectures, we eventually settled on a classic data modelling technique which worked very effectively.
We were able to show that our algorithm picked the same candidates for interview as a recruiter did 9 times out of 10, and that it always picked the candidate they eventually hired.
With only a small investment, the client was able to prove the viability of their AI technology before investing heavily in building a full product. The client is now proceeding with the second phase of the project, which is to build out an MVP.
The founder of Electric Brain, Bradley Arsenault, was formerly the Head of Research at Sensibill Inc, a local Toronto based startup.
When Bradley started working with Sensibill, they were a lowly startup with just the original two founders, a sales oriented CEO and a tech oriented CTO. Their technology for understanding receipts was primitive and based on hard-coded logic.
The basic problem for Sensibill was to take an image or an email of a receipt and to convert it into structured, database data. This is something that has been done with computers for 20 years using OCR technology, but most other products were only able to reliably extract the merchant name, the date and the total.
Under Bradley’s guidance, Sensibill built out an annotator and started creating an annotated receipt dataset. Many different algorithms were tried during this period, from automatically inferred rules, through to decision trees, through to nearest-neighbor algorithm over a set of hand crafted features, and eventually onto deep learning. With each iteration, the algorithm improved in accuracy and its resilience to errors, eventually achieved an average of 97.4% character level accuracy.
Based on this data-extraction technology, Sensibill went on to become fast growing startup and a world leader in receipt technology built for banks. Sensibill recently raised $17 million from various venture capital firms.