Bleenco Pitts

A framework for real-time predictive analysis on IoT devices.

The latest wave of AI-empowered products and services has made some remarkable achievements, especially deep learning in combination with audio and video processing. These AI engines run in data centers that were designed to scale automatically at low cost when needed – developers submit tasks via simple online APIs and receive predictions as an output – a true black box.
These AI black boxes perform fairly well when having a luxury of unlimited processing power and memory, but in-vehicle AI seeks for real-time predictive analysis on commodity hardware with no internet connection assured whatsoever. These requirements aim for specialized solutions designed explicitly for embedded platforms. Meet Pitts.

Data Preparation

Understanding your data and asking the right questions is a key to the successful learning and predicting patterns. Often times the data is not represented with simple numbers that are easy for machines to understand, but rather complex streams of data that require a substantial pre-processing.

Our approach is to parallelize as many tasks as possible to improve the speed, reduce the amount of data for complex processing to relieve the bandwidth and use hardware accelerated digital signal processing when possible to denoise the signal.

Learning Patterns

The amount of the data that is necessary for machines to learn patterns is enormous and therefore in many cases it may take weeks of machine training to produce first results. This is why learning processes usually take place offline, when there is plenty of resources available and a great variety of experiments can be performed to benchmark the outcome.

The outcome of a learning process is the model that describes how to separate learned patterns. This model is then one of the most valuable IP assets that will enable you to separate your organization from the rest of competitors in the industry that may use cheap generic models and online APIs.

Interested in learning more how can we help in protecting and benchmarking your AI models?

Real-time predictive analysis

The process of predicting results out of a trained model is called inference, which takes new data samples – that have not been seen earlier – as an input, shuffles them through the model and returns predictions. The real-time inference on embedded platforms cannot rely on online APIs that require internet access, models trained with the random data that was collected to answer different questions and inefficient software that does not allow to scale and improve.

We are hard at working on new algorithms and optimization methods to make embedded predictive analysis as efficient as possible. The 90% of our focus is on implementing and micro-improving the state-of-the-art algorithms. Only 10% of our time goes into inventing completely new concepts that might prove to be successful over next few years. Interested in learning more about our real-time solutions?

Interested in learning more about our real-time solutions?