Hello everyone, this is Michael, the head of BroutonLab. In this article, I would like to talk about how we helped the startup SeedMetrics, which specializes in the seed producer market in Brazil, to grow. This article will be of interest to founders and entrepreneurs who are building successful businesses by introducing technological solutions to the market. I will discuss how this startup emerged, and how we eventually transformed outdated processes for assessing crop yields for corn seed producers. At the end, information will be provided about the difficulties we faced, as well as the conclusions I drew. So, let's get started!
My future co-founders were looking for a technological partner ready to invest with them in a tech product that was still at the idea stage. They needed a strong technical team specialized in creating AI products. Additionally, they were considering sharing a stake in the startup in exchange for investment and strategic support. They approached our team to discuss possible collaboration.
The target audience for this product is seed producers. A distinctive feature of this market is the presence of labor-intensive processes aimed at obtaining accurate yield estimates. The significance of this problem is underscored by the fact that if fewer grains are harvested than stipulated in the contract, the producer is somehow penalized. Such problematic situations often occur due to the very nature of the seed plants themselves, which are weak, unevenly productive and susceptible to droughts, diseases, etc. Early prediction of these issues allows producers to take timely measures.
Since we specialize in creating AI products that, among other things, help to digitize data flows and business processes, making them transparent and more efficient, I was very interested in this problem. In this situation, I was also impressed that my potential partners had a deep understanding of the target market. To be honest, I know little about seed production, and developing this product myself would have been very difficult.
After discussing with my team, I decided that we would join this startup as technological partners. For a stake in the company, we provided:
- An optimal technological solution to this problem, based on our experience.
- Investments for product development.
- A team of developers who worked exclusively on this startup.
- The brainstorms we had were also very important for business structuration.
Let me tell you more about the problem we set out to solve.
Problem Being Solved
Corn seed producers regularly perform the following procedure to assess crop yield at early stages of maturation. Samples of ears of corn are taken in the field. These collected ears are then cleaned of leaves. Finally, a manual count of kernels on each ear is conducted. It's important to note that only healthy kernels are counted; diseased kernels or those too small must not be included.
To ensure kernels are not counted multiple times, the examiner often marks the counted kernels with a marker. The number of kernels for each individual ear is recorded in a notebook. This procedure is repeated many times. As a result, having information about the number of kernels on many ears, an average number of kernels per ear is calculated.
Now, to estimate the entire crop yield, the following quantities need to be multiplied:
- The size of the field (a known quantity)
- The number of ears per unit area (a known quantity)
- The average number of kernels (counted manually)
This procedure is carried out weekly during maturation. To perform it, a group of several people is usually involved. The number of people involved linearly depends on the size of the field.
Let's consider the disadvantages of this approach:
- People may rush and estimate the number of kernels by eye (as we later found out in pilot projects with our clients, the scale of this problem was even bigger than we expected).
- Subjectiveness when counting the small kernels: one person is always more optimistic than their peer.
- Classifying healthy kernels from diseased ones is not always straightforward. Moreover, I can look at a photo of corn for a long time, but even if I spend a lot of time, it's not certain that I will classify correctly! Can you?
As can be seen from the figure, there are diseased kernels that do not differ in shape from healthy ones, but they may have a slightly different color. Due to the need to review many ears in a short time, this problem is often formally resolved, leading to incorrect estimates.
- Recording observations in a journal leads to the loss of data over time, making it impossible to analyze retrospectively.
After analyzing the subject area, our team designed and developed the following solution:
- A Computer Vision pipeline, consisting of several neural networks and machine learning models. Essentially, these are the "brains" of our solution.
- A mobile application that allows for quickly photographing corn ears from different angles and sending them to the cloud for analysis.
- Based on AWS, we developed an energy-efficient and high-performance backend for data collection, analysis, and presenting the results in the form of analytics.
- A client's backoffice that includes role-based access. Its purpose is to assist in managing our solution within the company and to view analytical results.
It's worth noting that we made a number of important architectural decisions at the very start of the project, which helped us save a lot of money and nerves in the future.
- We used retool framework for database administration and to build dashboards for internal use. This no-code solution allows creating a page with all the important analytics of data received from our clients within a few hours. We completely abandoned the creation of custom admin panels on Django, which require constant maintenance and hosting.
- We moved all resource-intensive neural network computations to AWS Lambda. This allowed us to save money on servers and made our solution scalable and fast. No matter how many users we have, we can process many computationally expensive requests in parallel.
- The development of a mobile application on Flutter. This native framework is supported on all major mobile platforms. From our experience, it is much more convenient than React Native. Just our personal preference.
- We created cloud storage, where data can be stored in locations requested by the client. Moreover, one client's data is guaranteed to be isolated from other clients. This is a requirement that is important for our corporate clients with high data storage standards. If we hadn't considered it at an early stage, we would have had to rewrite the entire backend, which would have taken at least 2-3 months of development.
The development of the entire MVP took approximately 4-5 months. Refinement of our product, according to user feedback, is carried out regularly throughout the existence of the startup.
Research and Development
In telling the story of this startup, I cannot omit mentioning briefly about the R&D part. The main research part of our startup was to make an accurate prediction of the number of kernels in a corn ear based on several of its images. Our algorithm should not depend on the number of shots of the object. For example, our clients should be able to estimate the number of kernels from a single shot (on which there are several ears of corn). Alternatively, if they are willing to spend a little more time and photograph the ear from several angles, then our solution should process all of them and provide a more accurate estimate per ear.
Since the article is getting quite long, I've decided that our research, including the chosen models and architecture of the solution, deserve a separate article. After some time, I will write an article about how we developed the AI, the architectural decisions we made, how we built the data collection process from clients, etc.
I'll jump ahead and say that we managed to achieve 99% accuracy and our current clients, after conducting pilot experiments, reported that our solution works more accurately and quickly than a human. We were very pleased about this! 🙂
Challenges We Faced
Despite our product already being used by several large corporations, and me seeing great potential for development, our story hasn’t been straightforward. In this section, I would like to share the challenges we encountered while developing this product. So, let's get started.
Seasonality. One of the problems we faced was that our initial solution was only needed during the crop maturation process. Unfortunately, the laws of nature dictate that crops do not mature year-round. This situation complicates feedback gathering from clients and prolongs business iterations. Moreover, you only have one chance to make an impression. If we hadn’t immediately resolved the issues that arose during our first pilot project, at best we would have been offered to try another pilot in half a year during the next crop maturation. In the worst case, we would have been told 'thank you, we’ll think about it' and never heard back.
Data Bias. A peculiarity of creating AI-powered technology (for example, Computer Vision systems based on deep neural networks, as in our case) is data bias. During the MVP development, we did not have access to the real corn ears analyzed by our potential clients. Therefore, to build the AI system prototype, we used corn ears bought in a store and images from the internet. According to our understanding, we should have been able to process such ears well:
Unfortunately, in the real world, things were not so rosy. At the beginning of our pilot projects, our systems performed poorly. By the way, I never knew that corn could be red!
These kernels are covered with a red syrup that contains fungicides and insecticides. They are orange underneath, but our app must also be able to process such cases.
Not to mention artificial intelligence, which had never seen such images. And how about this case?
Or this one?
It turned out there are special hybrids of corn that don’t look as appetizing, but their genetics make them very good parents, and their kernels are more suitable for seeding. Consequently, we needed to quickly collect data from clients and retrain all our models, making them better and better. Again, a delay of 1-2 weeks meant failing the pilot project for the entire season and setting us back six months in our relationship with our partners.
Bureaucracy and slow processes at client companies. Since our main clients are global corporations, it takes a lot of time to communicate the value our product offers. This is a characteristic of the B2B market and in our case, it was distinctly pronounced. For instance, in one company, we have already conducted a successful pilot for four months, but we are still waiting for a decision from top managers about implementing our product.
Financial crisis. Currently, large agro companies are operating under tight budgeting due to the ongoing crisis. Large companies are restructuring, which manifests in the firing of top managers and the dissolution of entire departments. We have already encountered a situation where some managers, even knowing that our solution would help the company's business, were reluctant to take further steps, fearing to lose their job, prioritizing to save their jobs.
In summary, for innovative startups operating in conservative markets like agribusiness, the current time is not ideal. It's unclear when the ideal time will come. I’m not sure that sitting back and doing nothing is the optimal strategy.
Conclusions I've Made
From my experience, I've made several conclusions.
- I've once again caught myself thinking that flexibility within the entire team is very crucial in startup work. In this respect, I am very grateful to the development team with whom we have been working together for a long time. It is impossible to build a successful product if developers approach the work process formally. There are many unknowns in a startup, and bureaucracy usually does harm.
- It's very important to be able to choose partners. Do not work with people who are not on the same wavelength as you or whose values differ from yours. I doubt we would have achieved our current success if there had been disagreements.
- Maximize the speed of response to requests from your first clients. These are the people who help make your product useful. Thank them for this by striving to help solve their problems as effectively as possible. This will help you build trust with your clients and, as a result, help make a unique and useful product for the market.
- You need to be patient. In B2B, the decision-making speed is based on rationality. Having the best product on the market, you must be prepared that your clients may take a decision to purchase in 6 months or later. Strive to help your clients and keep in touch with them constantly. Especially if there are not so many of them in the world 🙂
Thank you for reading this article. I hope it was useful to you.
Who We Are
The BroutonLab team helps entrepreneurs and bootstrapped founders develop innovative AI-first products. We offer our partners our know-how, resources, and financial investments. Our mission is to develop innovative AI products and help bring them to market.
By the way, I use my own product and am working on several startups. If you have ideas for collaboration or any questions, I would be happy to answer them.
To stay updated:
- Add me on LinkedIn. I often publish thoughts on the development of technology startups.
- Subscribe to our blog.
Thank you all for your attention!