Robert Luciani
Aug 18, 2020

Using AI to Make an Impact on the Core Business


Narrowing Down the Scope of Ideas

The application of sophisticated AI to common use-cases is becoming increasingly commoditized, with cloud providers and firms providing clever APIs for entire line-of-business apps. When it comes to enterprise-readiness, the field of data science has certainly hit a new stride in the past years. Yet for all the available tools and “best practice” on how to deploy advanced AI workloads, many firms still find it challenging to move ahead on initiatives that have a notable impact on revenue.

In supply-chain and logistics it’s typical for more than 70% of operational expenses to be tied directly to the transport of parcels. Even so, data science teams in this vertical are often tasked with training extremely sophisticated and expensive NLP (language) models for use in call-centers. Statisticians are asked to set up linear models to do demand forecasting. These are the types of cookie cutter use cases that are presented at conferences because they apply equally well to all businesses, but by extension they tend to have a shallow impact. So how should a logistics company use AI to make a significant impact on its core business? The answer begins with providing the people in touch with day-to-day operations with technology that allows for ideas to be tested at the speed of curiosity.

Speed is Key to Creativity

If there is one commonality across the successful AI projects I’ve been involved in, it’s that they’ve all invested a significant amount of time on refining ideas and planning. At Foxrane we develop tailored AI models and high-performance algorithms for very niche business cases, but what often comes as a surprise to customers is how little time is needed for actual programming compared to other things. In fact, most of our time is spent on preparing qualitative and quantitative data for the project. To be clear, I’m not talking about mere technical matters of cleaning and transforming data, nor validating statistical priors or that sort of thing. I’m talking about a creative process where one begins with a business-model and aims to build a data-model that matches it. Remember, training an AI to make accurate predictions is not an end in itself but only a steppingstone. What additional data is needed to automate the entire business process in question? What data should be used to measure the effects of this automation? To this end OmniSci has become an indispensable tool already in the early stages of AI initiatives.

It’s a pithy saying that “speed isn’t everything” because it’s only true when you’re already fast enough. Imagine trying to write music on a guitar where every time you strum a chord the sound comes back after many seconds, or minutes, or hours – just like your typical business intelligence tools. I’d like to walk you through how we approach AI initiatives when the processing and visualization of enormous datasets is practically instantaneous.

Core Business Challenges Are Nontrivial

Let us consider a Swedish logistics business. In the animation below we see roadway and infrastructure information for the whole country combined with streaming fleet and courier IoT data, all overlayed on a single dashboard and rendered in milliseconds. In this particular case I’ve filtered 20 trucks that but have been standing still the longest and marked them red. They’re all at a distribution-center because too many vehicles were scheduled to load parcels at the same time.

Figure 1 - Good luck rendering this much data instantly on traditional GIS or BI software!

Now, you if you ask a statistician what to do about the trucks, they might excitedly propose modeling them using a Gaussian process to predict delays. However, I would suggest taking a step back to consider what the endgame is. The core of your logistics offering isn’t rooted in predicting delays. It’s probably something more akin to, say, providing same-day delivery with very precise time-windows. So, how do we avoid another “AI PoC” and use the totality of information collected in one place to solve a core business challenge? Let’s walk through a concrete example.

In the image below we see a scenario where a truck has become inoperable early in the day from a broken front window. The completed route is marked in green and the remaining route is marked in red. What we would like to do at this point, is see if we can maintain service level agreements by rerouting remaining trucks to cover the area marked in red.

Figure 2 – This image makes use of OmniSci’s native ability to render geometric objects. Roads, buildings, and other types of geospatial data can easily be processed and combined with relational tables and other traditional data sets.

Calculating an optimal route for a very large combination of vehicles, parcels, customers, time parameters, and more, can take weeks or even months using industrial solvers. Instead, what modern logistics programs tend to do is rely on expertly tuned heuristics to find a “good enough” solution. Something like a particle swarm optimizer can make complicated plans in just a few hours. We will take it a step further and train a Graph Neural Network to produce nearly optimal plans in milliseconds.

With such an AI in place, we can programmatically query OmniSci to filter and aggregate the data required for route planning at a moment’s notice. The result is a newfound ability to dynamically alter route plans and push them out to the LOB app carried by drivers. In the image below we see that in our admittedly trivial example, three trucks destined to handle nearby areas were dynamically rerouted and dispatched to handle the deliveries on the same day with only slight delays.

Figure 3 – Each real-life constraint such as delivery time windows, fuel limits, driver fatigue, and more increase the complexity of a plan exponentially. This is where AI really shines and produces approximately “good enough” solutions instantly.


Summary

Now granted, this example is an unconventional application of emerging AI techniques, but that’s the point. Well established solutions to general problems have a shallow impact, which is not to say they’re not worth pursuing. However, to achieve a significant and measurable impact on the core of your business, you need to plan further and be more creative than average.

What great innovators have known for a long time is that good ideas are very seldom a spontaneous gift from the muses. They’re born from a continuous and structured process of refinement. If you can afford yourself the ability to ideate at the speed of curiosity, you’ll find that great ideas materialize by themselves.

If you want to discuss an idea or simply find out more about the work we’ve done with OmniSci, Deep Neural Networks, and High Performance Computing, feel free to contact us at team@foxrane.com or request a time to speak with us here.

See you in the ether!

~ Robert Luciani


About Foxrane:

Foxrane is a software development firm focused on commercial applications of AI, HPC, and numeric analysis. Foxrane specialize in business-centric, cutting-edge technology to improve operating margins and unlock new revenue streams.