In fragmented industries, data is heavily siloed. Organizations are disincentivized from sharing access to first party data, diminishing its value and preventing meaningful insights that can only be derived from enormous datasets. We aim to ultimately address this issue across industries, but we have to start somewhere.
The annual global elevator service market is valued at $35 billion. Currently, the elevator service industry is fragmented into many small Independent Service Providers (ISP’s) and dominated by the larger Original Elevator Manufacturers (OEM’s) who possess troves of data. There has been increased investment in IoT technology in the elevator industry across all participants, but especially by leading OEM elevator companies Otis, Schindler, and Kone. Elevator service providers of all sizes are urgently seeking to leverage the data at their disposal (elevator performance and maintenance data) to develop predictive algorithms that optimize their service business models. In our interview, an executive at Otis noted that roughly 80% of their annual revenue comes from elevator service, so this is where their focus is.
Through Machine Learning and AI, predictive maintenance algorithms are emerging at the forefront of innovative elevator service technology. ISP’s are eagerly looking to leverage their data, but they don't have access to the same scale as the leading OEM's. Only the largest ISP's with more than 3,000 elevators in their network are able to develop meaningful machine learning algorithms. In comparison, Otis services 2M elevators annually. Datasets of that size enable neural networks and deep learning, creating a competitive advantage that could only be leveled through decentralized collaboration between ISP's on a peer-to-peer (P2P) network.
The blockchain technology outlined here is poised to disrupt the elevator service industry, but its potential extends far beyond elevators.
What it does
Our solution provides a decentralized P2P data sharing ecosystem for the elevator service industry that unites disparate data streams and makes it possible for smaller organizations to access predictive algorithms that are only possible with large pools of data.
Like the smaller players in any fragmented industry, ISP’s have been unable to leverage data to the same extent as larger OEM’s. Additional factors specific to the elevator service industry heighten its need for a trustless P2P data-sharing infrastructure, making it an ideal starting point on the path towards building a boundless data ecosystem for decentralized AI models. Disparate and unorganized data streams within the elevator industry create an immense administrative burden that further stifles the collaborative use of data to advance industry practices. Elevator service providers, though willing, are currently unable to monetize their data at all, let alone in an anonymous yet transparent way. Existing data storage solutions used by ISP's and OEM's also lack data immutability and accurate deduplication, making it difficult to train meaningful predictive algorithms.
The first implementation of this technology will be in the elevator service industry. Full Moon Ai will allow elevator service providers to store data from an elevator's controller and technician mobile apps on the blockchain. We retrieve data from an elevator's controller by integrating with technology like Modusystem that provides seamless access to elevator performance data. This data is aggregated with elevator maintenance data, which will be supplied via API integration with existing technician mobile apps.
By creating an industry-wide pool of both performance and maintenance data, we are able find insights that have never before been possible for ISPs. Drawing inspiration from centralized solutions in the industrial machinery industry (e.g. advancedtech.com)that use neural networks to perform oil analysis, thermography testing, ultrasound leak detection, and vibration analysis, we find that the possibilities for predictive models within the elevator industry are bountiful. We can, for example, create a predictive maintenance algorithm that predicts when specific elevator components will malfunction. It would also be possible to optimize elevator movement patterns to reduce energy consumption. Having talked to an executive at Otis, a vice president at an ISP with a network of 4,000 elevators, and a building owner, these seem to be the most addressable needs in the existing elevator service industry. As such, this is where we plan to start.
How we built it
The foundation of our solution lies on the decentralized blockchain infrastructure that creates a secure and transparent system for anonymously sharing data and accessing AI models. With no single point of failure, the distributed ledger that immutably tracks all transactions on the network can enjoy 100% uptime. Several cryptographic primitives such as the SHA256 hash function and Zero-Knowledge Proofs enable the anonymity necessary for handling highly sensitive information.
Just as Snips AIR (snips.ai) uses their native cryptocurrency to incentivize users to add their audio data to the blockchain for training their AI algorithm, we will similarly incentivize ISPs to connect their elevator controller and maintenance data to our platform. Predictive algorithms like the predictive maintenance model we plan to offer will live on the blockchain as Smart Contracts. These predictive algorithms will be accessible by everyone, but the ability to commit data to the blockchain will be strictly permissioned; only organizations with the proper authorization will be able to add their elevator controller and maintenance data. Whenever someone (ISP, OEM, Building Owner, etc.) utilizes a predictive algorithm on the network, the Smart Contract debits their account in the native cryptocurrency. This amount is then used to compensate the data suppliers based on their relative contribution to the training dataset of that specific algorithm. A fixed Developer Fee will be included to compensate the developer who identified the relationship within the data and published the Smart Contract. This fee is fixed to avoid price undercutting in the market.
During launch, the ability to add Smart Contracts and commit data to the blockchain will be solely administered by Full Moon Ai in a centralized manner to further discourage bad actors and allow us to develop the first predictive algorithms for the elevator industry. In the future, we aim to allow independent developers to create their own Smart Contracts on the blockchain, but only after we solidify our foothold in each industry. ISP's, Building Owners and OEM's will apply for access to the network and provide necessary information such as an elevator's ID numbers and the locations of each elevator. Hashed versions of this data will be stored on the public blockchain, and as such, the actual data will remain concealed from others. In the future we aim to move towards true decentralization of governance, and having recently spoken about this with Alex Christian at DataMynt, we are reassured that it is a viable option to incrementally work towards this goal.
Challenges we ran into
We desperately want to be able to demonstrate value to our audience before we build the V1 implementation. This is a challenge for us because, as we learned from our customer interviews (discussed below), OEM's and ISP's are reluctant to share access to their data without compensation. Our original goal for this week was to take two datasets and draw a meaningful conclusions using R. We are still working towards this goal, but will likely need money to compensate data providers.
Another challenge is identifying the optimal Consensus Protocol for our implementation. We reviewed dozens of Whitepapers and academic studies that indicate that Proof of Authority or Proof of Importance might be viable options that provide proper throughput on the network. These Consensus Protocols, however, compromise on decentralization of nodes in favor of increased scalability. This tradeoff, inherent with the scalability trilemma that faces all Decentralized Applications (Dapps), must be further evaluated for our use case.
As two non-technical founders, most people like us would be discouraged from tackling such a complex problem. However, having founded and scaled a 22-person full-service healthcare marketing agency that works with some of the nation's largest private medical practices to create digital experiences and online marketing campaigns, we have the skills necessary to recruit and lead a highly technical team. We are currently collaborating with two Duke seniors who plan to join us on our journey once we return to campus in the fall. These two additions to our team will bring experience from Facebook, IBM, and Bridgewater Associates that will be helpful in designing our token economics model and building truly decentralized AI algorithms.
Accomplishments that we're proud of
At the beginning of the week, we had yet to talk to any of our stakeholders in the elevator service industry. We quickly understood that developing connections with OEM's, ISP's, and building owners would be instrumental in narrowing the focus of our first implementation, getting real-time feedback, and ultimately piloting our solution.
To do this we first signed up for a free account on Hunter.io, a software we identified that easily finds names and emails associated with a specific website. We compiled a list of ISP's and their websites, and then uploaded each to Hunter.io to retrieve the emails for outreach. We also wanted to speak with building owners and executives at an OEM, but entering a website for these audiences on Hunter.io would return too many emails for us to find the most appropriate contact. We then found GetProspect.com, a Google Chrome extension that finds email addresses associated with specific LinkedIn profiles. We then meticulously searched for the right contacts on LinkedIn and began reaching out.
We performed one round of outreach to a total of 58 contacts. We have scheduled a follow up email that will go out next week, which we expect will increase our response rate. Of the 58 emails sent, 7 bounced (email address was invalid). We received 8 positive responses with people willing to help in any way they can. We managed to quickly schedule 3 calls. We spoke with an Otis executive, a large Manhattan-based ISP, and a building owner. In the process, we received 4 warm introductions to other stakeholders in the industry and sent 6 additional surveys to a mixture of ISP's and building owners who we have yet to speak with over the phone. We have 2 more phone calls with ISP's scheduled for next week.
What we learned
Immediately after sending our first round of cold email outreach we began receiving positive responses from people from all types of organizations who are more than willing to assist in our research. Our hypothesis is that people are generally more willing to help students, so we strategically wrote our outreach template in casual tone and kept the email short and sweet. We focused mainly on who we are and why we'd like to talk to them. Here is an example outreach email we sent to an ISP:
My name is Andrew and I’m a rising Senior at Duke. The reason I’m emailing is because I’ve been working with Duke engineers to create a solution in the elevator maintenance space that leverages data and AI. I would really love the opportunity to quickly chat with someone in the community who either deals with elevator management or maintenance and I thought you might be someone I should ask.
Would you be willing to answer a couple questions for me or refer me to someone who might? Much appreciated.
Something else we learned is that the Duke community extends far beyond Durham. We had a chance to talk with Aaron Udler, a Fuqua alumnus who introduced us to Alex Christain. Alex Christain leads the team at DataMynt, a venture backed DeFi solution, and he also attended Fuqua. These conversations were invaluable and helped us identify dozens of Whitepapers that we subsequently tore apart. As a non-techncial founder, Alex also suggested several resources that he uses to help validate his technical decisions. These will be useful as we write our own Whitepaper.
What's next for Full Moon Ai
We intend to pilot our solution by partnering with 2 ISPs to aggregate their data and create a predictive maintenance algorithm as a Smart Contract on the blockchain. Our long-term vision for predictive maintenance is to make part-specific predictions, but since this won't be possible with just 2 ISPs, we will limit ourselves to a handful of data points to reduce the amount of time needed for training the model. One way that we will do this is by categorizing the parts into groups to make category-specific predictions. If we achieve this, it will demonstrate that even with limited data, our solution still demonstrates a significant improvement from existing predictive capabilities in the elevator service industry. All existing solutions that provide predictive maintenance algorithms for elevators (Uptime, LiftAI, LiftInsight, etc.) are unable to make part-specific or category-specific predictions, so this would be a sizable advancement that would cause problems for these companies and pose a threat to OEM's primary business model.