Moving Forward with AI: Ritesh Ratti from Delivery Hero on AI Advancements in Logistics

Logistics is moving forward at its own pace, testing the latest technological advancements and adapting to AI progress. While generative artificial intelligence offers more possibilities in other niches, logistics is also trying to harness it for its own needs.

In a conversation with Ritesh Ratti, Jarek Jarzębowski asks about key issues at the intersection of AI and logistics. What are the biggest challenges of implementing large language models in this sector? Are pre-built LLMs and Retrieval-Augmented Generation useful for logistics? How can a logistics company set its data science journey in the right direction?

Find answers in this insightful interview, where Ritesh Ratti shares his industry knowledge.

Key Takeaways from the Conversation

AI has put logistics on a direct route to efficiency: Instead of building manual rule-based models, logistics companies can now implement models based on labeled datasets, saving resources and expanding productivity.

Implementing AI in logistics can be a rocky path: Effective implementation remains the biggest challenge, as large language models can be time-consuming to process. Careful planning and tool selection are essential for optimized AI deployment.

The optimal route for LLMs in logistics is through fine-tuning: Using pre-built large language models can be risky in logistics, but fine-tuning techniques and reinforcement learning with human feedback will keep your solution on the right track.

Generative AI can set a new direction for various areas of logistics: Automation of the supply chain, product sizing, and anomaly detection are where Generative AI shows the greatest potential in this industry.

Conversation with Ritesh Ratti:

Jarek Jarzębowski: Can we start with you telling a little bit about yourself, your experience, and your background, especially focusing on using ML, AI, data, and logistics?

Ritesh Ratti: Sure. I’m Ritesh Ratti, currently working with Delivery Hero as the data science lead. I’ve built projects related to food taxonomy generation, using language models to develop food categorization and sub-categorization.

 I have over 13 years of experience, starting as an engineer and moving into data science and machine learning. My career began at Samsung, developing a question-answering system, followed by work on recommendation systems at Aerial and Pitney Bowes. At Grab, I focused on recommender systems and food categorization using transformer-based modeling. 

Currently, I lead a team of data scientists and engineers, working from POC stage to production. I hold a doctorate in data science and a master’s in computer science, with a focus on NLP-based models for anomaly detection, specifically on network data.

Jarek Jarzębowski: Great. Most of our readers are familiar with AI, data, and machine learning, but for the general public, AI has become more common with generative AI and large language models. Can you give us an overview of the AI landscape, especially in logistics?

Ritesh Ratti: The landscape has changed significantly over the years. Ten years ago, we relied on rule-based models for specific problems like address parsing in logistics. We would manually extract rules to parse addresses into different components, like street, postcode, and building name.

Now, machine learning allows us to build models based on labeled datasets, shifting away from manual rule-based systems. Generative AI is used to extract meaningful information from text, like addresses, using large language models trained on multilingual data. AI has also advanced in identifying products through images, which helps in estimating shipment costs more accurately. 

In logistics, AI plays a vital role in computer vision models, product classification, and even optimizing models for production using technologies like Kubernetes and C++/C-based frameworks.

Jarek Jarzębowski: Can you explain the biggest challenges in introducing these models and getting them into production?

Ritesh Ratti: The biggest challenges are scaling these models to predict outcomes quickly and efficiently and then deploying them in production. Large language models can be time-consuming to process, so we use solutions like Kubernetes to distribute processing. We also optimize models through quantization, but we need to balance performance with accuracy. 

For production, we often use FastAPI or Flask to expose models as APIs, and we deploy them using Docker and Kubernetes. Other tools like Spinnaker help automate deployments. Overall, it’s a complex stack, and careful planning is needed to ensure models run efficiently in production.

Jarek Jarzębowski: How useful are pre-built large language models in logistics compared to those you need to build from scratch?

Ritesh Ratti: Pre-built large language models can be challenging because they may not have been trained on data specific to your needs. However, you can use techniques like prompt-based engineering or few-shot learning to adapt these models by generating data to fine-tune them. 

Reinforcement learning with human feedback is another strategy to improve these models iteratively. The key is to balance the use of pre-built models with custom fine-tuning to meet specific requirements in logistics.

Jarek Jarzębowski: Are retrieval-augmented generation (RAG) models useful in logistics? Can you explain what RAGs are and how they work?

Ritesh Ratti: RAG stands for Retrieval-Augmented Generation. It involves using a language model and providing instances of information from a vector database to answer questions or generate content. 

RAG models first index data in a vector database for efficient retrieval and then use similar instances to generate better responses. This approach is beneficial in generating accurate solutions for complex problems by leveraging in-context learning, where the model uses provided examples to understand the context better and produce more relevant outputs.

Jarek Jarzębowski: With the rapid advancements in AI, do you see more opportunities for data science in logistics and supply chain in the near future?

Ritesh Ratti: Yes, there are many possibilities. Generative AI can enhance human interaction with systems, enabling deeper understanding and automation in tasks like product sizing and anomaly detection. 

AI can also be integrated into robotics for supply chain automation, reducing human intervention. Additionally, multimodal generative AI can help analyze issues reported by consumers, identifying whether there are genuine product problems or fraud attempts.

The potential for AI to streamline and improve supply chain operations is vast, especially with the continued development of generative models.

Jarek Jarzębowski: What advice would you give to companies in logistics that are just starting their data science journey?

Ritesh Ratti: First, identify the areas in your logistics pipeline that need improvement, such as product optimization or machine learning solutions. 

Then, assess the risks of implementing machine learning and have a backup plan in case results are not as expected. It’s essential to have a team of experts who understand your product and can build effective solutions.

 If using generative AI, consider its cost and performance impact, and start with simpler approaches like prompt engineering before diving into more complex fine-tuning. 

Lastly, ensure security measures are in place to protect user data, especially with GDPR compliance, and be cautious of potential issues like model hallucinations.

Jarek Jarzębowski: To have a successful product, we need experts in technology, security, and business. From your experience, which of these components is most often missing?

Ritesh Ratti: The most important aspect missing is business understanding. We need to evaluate from a business perspective whether we really need to invest heavily in machine learning, and to what extent, to ensure better user experience and customer satisfaction. 

A simple solution for a big problem might be the best approach. If predefined rules can solve the issue, that’s the direction we should take, especially if those rules are stable and not likely to change soon. 

Explainability is also crucial, especially when reverse engineering or identifying loopholes in models or data. These are key factors for driving the business forward.

Jarek Jarzębowski: One drawback of using data science is the inability to precisely predict outcomes and ROI beforehand, unlike traditional software development. How can we cope with these challenges when we don’t know if the solution will even be possible or how much it will cost?

Ritesh Ratti: That’s a great question. First, you need to check if historical data is available for the problem you’re trying to solve. Analyzing this data can help estimate ROI, like predicting increases in revenue if certain features are implemented. 

Data analytics plays a vital role here. You should also consider costs like infrastructure and resources, and whether you can reduce man-hours. These factors together can help estimate ROI.

Jarek Jarzębowski: Can you give us some predictions about the challenges that might arise with using more of these technologies in the future?

Ritesh Ratti: I see a few major challenges. One is resource availability, especially as trends move towards building proprietary models. 

Another is the environmental impact of training large language models, which can run for weeks and consume significant energy. Explainability remains crucial, especially in risk management, as incorrect predictions shouldn’t bring down an entire system. 

Security is also a concern, as providing false information could lead to disastrous consequences. Lastly, the impact on education is significant; while these models can enhance knowledge, misuse could hinder conceptual understanding.

Jarek Jarzębowski: There’s a lot of potential for good with technology, but we also need to consider these challenges. Thank you for sharing your perspective and experience. Is there anything else you’d like to mention, especially regarding the use of data science in logistics?

Ritesh Ratti: Yes, there are many areas in logistics where data science can have a big impact. Inventory management is one, where you can invest heavily in estimating inventory levels and locations. Optimizing delivery routes and ensuring on-time delivery are also crucial areas. Another aspect is verifying the authenticity of delivery addresses. There’s a lot of potential in the logistics and supply chain sector, and it requires more research. It’s a huge opportunity.

Jarek Jarzębowski: We might delve deeper into these topics in the future, but for now, I think we’ve covered a lot of ground. This will make for an interesting article. Thank you again!

Ritesh Ratti’s Background

Ritesh Ratti is a dynamic Data Science and Machine Learning Lead with over 14 years of experience in crafting intelligent software solutions. Leading successful teams at HelloFresh and Delivery Hero, he applied his expertise in machine learning, data mining, and predictive modeling to drive impactful insights. 

Ratti’s work is rooted in a deep understanding of classification, clustering, and text analytics. With a Ph.D. in Computer Science from the Indian Institute of Technology, Guwahati, he has also presented research at top conferences in machine learning and network security, pushing the boundaries of what’s possible in data science.

About Delivery Hero

Delivery Hero is a global leader in online food delivery and a pioneer in quick commerce, operating in around 70 countries across four continents. The company, headquartered in Berlin, was founded in 2011 and has rapidly grown to become one of the world’s largest online food delivery platforms, handling millions of orders daily. 

Delivery Hero not only connects customers with restaurants and grocery stores but also focuses on innovative technology solutions to enhance the delivery experience, including the use of AI and automation. Committed to sustainability, it actively works on reducing its carbon footprint across its global operations.

Closing Thoughts

The recent advancements in AI and GenAI may set a new direction for logistics in the next few years. From computer vision powered autonomous solutions in ports or manufacturing halls, through predictive maintenance, to simulating optimization scenarios, the logistics today is already strongly AI-driven. 

As Ritesh Ratti underlines, for these implementations to succeed, they need to be followed by careful planning and tool choice. Contrary to common conviction, even the pre-built large language models can benefit logistics with the right approach. We cannot wait to see how this new phase of AI’s development impacts and improves the industry!