A Conversation with Saurabh Goyal
Good morning, Saurabh, and welcome to Commodity Conversations. I am keen to learn more about AI’s impact on the supply chain for agricultural commodities.
Good morning, Jonathan. Thank you for inviting me to your platform. We are all in a learning phase, but I am happy to share everything we have learned.
First, please tell our readers about yourself. I see that you studied at the Indian Institute of Technology.
Yes, I did a course in mechanical engineering, after which I worked as a programmer with Tata Consulting Services before doing an MBA in marketing and finance. I then worked as a program manager for the New York Mellon Bank, where I developed some trading systems. Later, I worked for Accenture Capital Markets and Prudential Asset Management. In both positions, I worked in development and implementation roles, managing trading systems for financial instruments.
Olam International hired me in 2009 as their head of CTRM – Commodities Trading and Risk Management Systems – where I developed trading systems for physical commodities. I was with Olam for eight years, working in almost all aspects of commodity trading operations, including supply chain and risk management, hedging and trade finance. We developed many of their in-house operations systems for cocoa, cotton, coffee, sugar, and grains.
What made you leave Olam and start your own company?
As Olam grew, the company replaced their in-house systems with SAP. I was initially involved in SAP implementation, but I quickly realised that SAP was more complicated, expensive, and challenging to implement than our in-house systems. That realisation gave me confidence that there was a market for small and medium-sized commodity traders – like Olam 20 years ago – who could not afford to implement SAP or develop an in-house solution.
Some companies provide front-office applications for order and risk management. Still, few give end-to-end solutions, including the front office and hedging, operations, accounting, middle and back-office activities, trade, and finance.
We founded Phlo Systems in 2016 to provide an end-to-end, easy-to-use, fully custom-built solution – something built on-cloud for consumption on-cloud.
We do almost everything from order management to sustainability. Our latest project is to help companies manage the traceability required under the new EU deforestation rules.
We are around 30 people. Thirteen are in the UK. The remaining are in India, Ukraine, Bulgaria and Ghana, West Africa. Our smallest client pays $150 monthly, and our largest pays $15,000 monthly.
Do you work with Blockchain?
We were initially optimistic about Blockchain, but after one year of working on it, we realised it was not a solution and stopped.
Many people gave up on Blockchain after investing significant time and money. Blockchain was designed specifically for Bitcoin; it is the technology behind it. Our biggest mistake was to think we could separate Blockchain from cryptocurrency.
The only place where Blockchain will work is cryptocurrency. It is the only use case for it. Taking the technology that drives cryptocurrency and using it for something else does not work. It is like trying to fit a square peg in a round hole.
Do you use AI?
We began using AI three or four years ago. We used it for scanning documents like bills of lading, invoices, packing lists, and origin certificates and then auto-filling forms in the system. It saved someone manually filling them. We used an AI technology called Amazon Textract, but every time the format of a document changed, you had to make changes in the system setup – and it was not 100% reliable. It made mistakes in reading the data and importing it into the system. As a result, we were cautious about including traditional AI in our applications.
Chat GPT and Open AI have made things 100 times better. We now feed all our documents to Chat GPT through the back-end APIs (Application Programming Interfaces).
Is it more reliable?
It has so far provided 100% reliability and accuracy in reading the data from the documents and importing it into our application.
It is an example of how generative AI has completely changed the landscape. You previously had to put a lot of investment and effort into integrating traditional AI into your systems, but with Chat GPT, all that hard work is gone. You apply basic software engineering to integrate Chat GPT and ensure the process flow works fine. Chat GPT provides the intelligence to read, interpret, and structure the data, leaving us free to focus on the engineering part.
Some might say that our enthusiasm for AI is like what we initially had for Blockchain, but it isn’t. With Chat GPT, we are already reaping the benefits.
Will Chat GPT be transformative for our business?
I believe so. We are now looking at many different use cases where we can start using Open AI and similar models.
User support is one obvious application. No matter how wonderful your system is, you must provide client support, especially regarding ERP (Enterprise Resource Planning). Our scalability as an organisation was dependent on being able to hire resource people to give that service to our customers. Chat GPT can automate most client support functions, leaving human interaction only for critical use cases.
So, you are using AI for CRM (customer relationship management) and document processing. Anything else?
Absolutely! Much, much more. I just gave you examples of traditional use cases where Open AI is a game-changer because it does the job 100 times better.
But there are 100 other use cases which were earlier not even possible. For example, we have a customs module where an international trader can make a customs declaration without using a customs broker, an agent, or a freight forwarder.
So far, we have only done it for the UK. All the information you need to make a UK customs declaration is available on the HMRC website. We have extracted that information and trained Open AI to do the work. The trader explains the scenarios, for example, importing animal products from Germany to the UK, and the system guides them step by step. Our clients appreciate this kind of functionality.
Is it something that you could repeat in other countries?
We are working on it for the Netherlands. The module should be ready by the end of this year.
Is language a problem?
No, but it’s a good question. Chat GPT was trained only in English, and no one expected it to understand any other language. No one intended it to be fluent in different languages, but it read foreign language documents during the training process and taught itself.
We have done a proof of concept for a Polish freight forwarder who wants to use Chat GPT to process their documents, many in Polish and Italian. We thought the system would fail, but Chat GPT could interpret them 100% correctly.
Language from an AI perspective is not a problem. However, language from a user interface perspective is. We’ll have to change our system to display other languages, but it should not be a problem.
Okay, that’s a perfect example of how you’re implementing AI. Do you have other examples?
One useful plugin converts a natural language question into a structured one that the database understands.
In our application, for example, the user can ask a question like, “List my ten top customers last month, based on the number of transactions, but only the customers exporting from the UK into Germany.” It is a complicated query, and you can make it even more complex by saying that you’re looking for customers dealing with robusta coffee or organic grains. We have developed a plugin to take this query in English, map it to your database structure, convert the English query into SQL (Structured Query Language) and provide you with the answer. We are still in the proof-of-concept stage, but it is reasonably accurate.
Do you have to pay to use Chat GPT?
Chat GPT has a free-of-charge open interface and another with an internet connection and several 3rd party plugins that costs $20 per month. It gives better results.
The problem with the open-ended interface is that you interact with the Chat GPT model and the data within that model. You can’t do anything beyond that. In the customs example, Chat GPT will not function correctly as some customs information sits behind a firewall on the HMRC website, and you need to enter your username and password to get it.
Similarly, Chat GPT will not be able to access your in-house database. However, Open AI offers you back-end APIs through which you can connect your data and allow access.
Are you worried about allowing Chat GPT access to your data? Isn’t there a data protection issue here?
These large language models allow you to protect your IP (Intellectual Property). You don’t provide Chat GP with raw data; you first convert it into vectors – a matrix of numbers.
When you ask Chat GPT a question, it converts it from text format into a vector using “embeddings”. Embeddings are the science behind all these language models. Chat GPT uses nearly 20 billion parameters to convert a question into a vector. It then compares it to all the other vectors it has stored, looking for the closest match and coming up with an answer. Chat GPT doesn’t understand words; it understands the numbers and vectors created from words.
You can provide Chat GPT with your raw data to process, or you can convert it first into vectors. The Chat GPT model is open because it discloses the embeddings you need to convert your data. You can apply these embeddings to your data, transforming it into a vector format before sending it to Chat GPT. It then sends a vector back as a response. It doesn’t get to know the raw data behind the vector.
There may be a counterargument that Chat GPT can apply reverse engineering on the vector and return to the original data, but that process is computationally heavy. If Chat GPT started to do that, it would need an enormous amount of hardware. It’s not computationally feasible for them to do it.
I can’t imagine Chat GPT letting you do all that for £20 per month. Or are they?
No, the API connectivity charges are based on the number of tokens we send to it in each question, but the pricing is not exorbitant. We spend less than £500 per month as of now. It will be an issue if Open AI increases its pricing, but there are other large language models. Open AI is not the only game in the town. There are many different models, and some are free of cost. They are not as good as Open AI but are catching up.
Meta, the company behind Facebook, has released an open-source model and pledged never to turn it into a commercial venture. But then Open AI started as an open-source model before Microsoft converted it into a commercial model when they realised its massive potential.
You never know what’s going to happen in the future. Open AI is currently reaping all the benefits. But there is a consensus in the development community that an open-source model will prevail in the long run. And it won’t be Open AI.
How will AI transform the agricultural commodity trading business?
It is a difficult question to answer. As software developers, we see that it is transformative in developing and providing our systems to our users. GPT is excellent with the English language, but it is ten times better with software programming languages.
Microsoft owns Open AI and GitHub, an online repository for code. We suspect that Microsoft trained Open AI on the software code stored on GitHub. It has made Open AI an excellent programmer. It generates perfect code in any language you want. It is also suitable for testing your code, finding bugs and issues, et cetera. It has made the software development process five or six times faster than without it. So, it is transformative for us as a software development business.
I believe that Chat GPT will transform the way users consume software. Users currently consume software in a form-like interface, entering and saving the information. It’s a data-driven approach.
In a few years, we will see conversation-driven user interfaces. You will tell the system what you want to do. For example, you will tell it that you have received a new contract to supply a commodity to a customer. It is lying in my email. Can you please check and store it in your system?” And the system would be able to do that.
Will AI make some professions redundant?
The politically correct answer is, “No, it won’t. It will help professionals do their jobs better, remove the mundane tasks and empower workers to focus on higher-level thinking.”
The correct answer, however, is, “Yes, AI will make some professions redundant. We’re already seeing it. Shares in Chegg, the biggest tuition-assistant company in the US, lost more than 40% within six months of Chat GPT launching. The company connected kids to different teachers, helping them with their homework, but Chat GPT does a better job than many teachers.
UK schools have banned children from using Chat GPT. I think it is a shame. You can’t stop technology; you must use it correctly rather than prohibiting it.
Chat GPT was supposed to be a purely language model; it was not supposed to solve mathematical questions. However, it solves mathematical questions by breaking down any question into logical steps. Some people still think Chat GPT is just a stochastic parrot, but it is not. It is much more than that.
So, Chat GP wasn’t supposed to understand non-English languages or solve mathematical problems, but it does. Were there any other surprises?
Yes, but they were not all positive; some were negative. Hallucination is the most negative. It is when Chat GP gives you a wrong answer with confidence. No one expected it to do that, but it does.
Should we be afraid of AI?
I don’t think so. AI doesn’t have an agency; it will not suddenly wake up at night and start thinking for itself. It is a program residing on a system, waiting for you to ask it a question. When you ask a question, it looks back into all its learning and provides an answer. It is not fundamentally designed to think for itself.
The doomsday scenarios are not applicable here. What is possible is a wrong actor will use it for destructive purposes. Safeguards are in place, but you can trick Chat GPT into answering a question it has been told not to answer.
What is the difference between AI and machine learning programs?
With machine learning, you train a machine for specific data and context-driven tasks. AI is not dependent on any particular use case. A machine learning program trained to do one thing couldn’t teach itself languages or maths.
Could AI learn to trade on the futures markets for speculating or hedging?
Electronic and high-frequency trading, identifying micro patterns and then acting on those patterns, have been in existence for many years. Companies have made fortunes using the technology. But their success has depended not only on their mathematics and algorithms but also on their better connectivity to the exchanges and some information-based edge.
Trading companies tend not to try to identify ultra-short-term trends but look at a broader horizon based on supply and demand imbalances.
Chat GPT will not revolutionise short-term trading because short-term traders analyse real-time tick-by-tick data; it is not Chat GPT’s model. Short-term trading will continue to rely on statistical probabilistic models already out there. It’s a mature industry.
Chat GPT may be better suited to longer-term trading, looking at supply-demand imbalances – crop and weather reports and the like – and then taking a direction-based strategy. I think it can work.
Thank you, Saurabh, for your time and input.
© Commodity Conversations ® 2023