Getting physical to do more with small data
Getting physical to do more with small data
TUBR uses physics to predict next moves PLUS a startup update
Hello there,
If you’re into A.I. startups, and I know many PreSeed Now readers are, you’ll enjoy today’s profile. Scroll down to read all about how TUBR is using physics and machine learning to help SMBs make predictions about people, objects, and physical spaces.
We’ve also got an update on a startup we covered earlier in the year. I hope to do more of these in the future. Early-stage startups often evolve over time as they learn more about the market they’re serving or creating, so a check-in can provide valuable insight.
And you might have seen that Substack (the platform this newsletter uses) launched a chat function last week. I’m planning to experiment with it as a bonus for paying PreSeed Now members soon - stay tuned!
– Martin
🚨 Status update
Catching up with a startup we’ve previously covered:
Payful was one of the first startups PreSeed Now ever covered back in May. At the time it was trying to solve late-paid invoices with a community approach. As I wrote at the time:
Billed as “a community of businesses with good track records of paying invoices on time”, Payful integrates with accounting software to track how quickly businesses pay each other. This generates a ‘Payful score’ for each business, letting everyone else in the network see who is a reliable payer. And because each company knows they’re being judged on the promptness of their payments, there’s a strong incentive to maintain a good score.
Since then, they’ve learned from user feedback and have evolved their offering. Founder Andy Taylor writes:
“Our hypothesis that businesses wanted to know how their prospects paid their invoices through behavioural analysis was true but only for around 10% of the market (enterprise clients). We knew that in order to eliminate late payments we needed to make Payful available to 100% of businesses across the globe.
”Back to the drawing board we went and figured that what 90% of the market actually wants is to be connected to businesses that pay on time. At this time, we were about to start a pre-seed raise but decided to put this on hold until 2023 while we conceptualised our findings.”
With the original offering now aimed squarely at enterprise customers, Payful is building what it describes as “the most sophisticated B2B prospecting machine the world has ever seen” for the SMB market. Essentially, it’s a way of discovering companies with good histories of paying their invoices on time, verified by Xero data. Taylor continues:
“We're excited because the hand-in-hand both products can eliminate late payments for all businesses from freelancers to large enterprises across the globe.
”Within the last few weeks we've added almost 30 businesses to the waitlist and currently speaking with two large enterprise clients on how Payful could be used to help protect them from invoice fraud. We decided to open our round early and commenced pitching last week. We're hoping to close this by the end of the year.”
TUBR is going big on smart predictions from small data
‘Small data’ is a growing field that looks to bring the benefits of machine learning to businesses that lack large data sets. Gartner predicted last year that by 2025, “70% of organizations will shift their focus from big to small and wide data, providing more context for analytics and making artificial intelligence less data hungry.”
Sheffield-based TUBR (pronounced tube-er) is looking to carve out a unique niche in the space by leveraging expertise in physics to make predictions about physical space and objects to help customers make the most of their resources and optimise customer experiences.
Co-founder and CEO Dash Tabor says that small and medium-sized businesses often miss out on being able to use predictive technology because their data isn’t robust enough, they lack the know-how in their team, and lack the time needed to get a solid machine learning operation up and running.
TUBR wants to solve all three of these challenges. “Our model works with both complete and incomplete datasets, and can handle the chaos in the system. We handle all the data wrangling and make it super-simple and easy for them to pass us data and us return a prediction data set,” says Tabor.
“Normally it takes a lot of time to get companies up and running and building machine learning models. We say we can do it in less than a month, but to date it's taken about a week and a half. So we have a fast turnaround for people to start seeing value.”
TUBR’s specific focus is spatial problems - predicting movements of people or objects. For example, being able to predict within a 20-mile radius where a car is likely to break down, how many pies are going to be needed in different parts of a football stadium, or footfall rates in a train station with an eye on maximising retail opportunities in the space.
Let’s get physical
The technical chops behind TUBR come via the work of co-founder and CTO Nikhil Kanta, who has a background in both physics and machine learning and was looking for a way to bring the two fields together.
“Typically, machine learning is right now being done in a statistical approach, where you have a bunch of data, you try to find the trends in the data, and try to make predictions based off the trends,” says Kanta.
“What we are doing here at TUBR is to take the data and run it through what we call our ‘physics engine’. It uses data to reconstruct the system the data is from, help us explore the hidden variables in that system, build a model of that system, and put physical constraints on the machine learning to look at proper spaces to learn and to make more accurate predictions. Physics helps us understand the system more than just understanding the trends and the data.”
Tabor says the kinds of SMBs that work with TUBR at its current early stage usually want to improve their operations by optimising their assets and reducing waste, with an eye on enhancing customer experience. These companies sometimes don’t know exactly how the data they have will help, but TUBR can assess it under NDA to figure out what will help with predictions.
“The physics engine finds the physical constraints, the hidden variables, the impact factors that help it understand what that data is doing and how the data points connect together and builds a model. Then we run the data through the machine learning and we're able to tell them ‘this is how accurate we can be for your objective’.
From there TUBR is able to deliver new predictions up every five minutes, although Tabor says most customers are happy with daily, half-daily, or hourly predictions. Predictions can be made up to 21 days in advance.
Growing with small data
To date, Tabor says TUBR has done a number of proof-of-concept projects with customer data, resulting in an 85% accuracy rate at first, rising above 90% based on continuous learning from the data.
Tabor says that a test involving open-source Transport for London data was able to predict lower footfall at Baker Street tube station on days when there was industrial action, based solely on changing patterns in the period leading up to it.
When I spoke to Tabor and Kanta, they were just about to increase the size of their team from five to six. The pair first met while taking part in an online accelerator programme called SeedReady at the start of the pandemic, when Tabor was looking for a way to launch a startup based on her extensive experience as a data product manager at companies like Experian, Streetbees, and Techstars-backed Tripcents.
Tabor says while everyone else told her the idea of TUBR was impossible, Kanta said he had a way to achieve it. A very modern startup, they’re an international cofounding team. Tabor splits her time between London and Sheffield, while Kanta is based in India, but the commercial team is based in, and will grow, in Sheffield.
Investment plans and future outlook
Keep reading with a 7-day free trial
Subscribe to