**Interested in This Technology?**

If you want to learn more or collaborate with us on this project, contact us at info@deix.ai.

Imagine having to solve a complex problem, such as organizing the optimal employee shift schedule in a company while minimizing costs and complying with all legal and productivity constraints. For an operations research expert, this means transforming the request into a mathematical model, with precise equations and constraints, and then delegating the solution to specialized software.

But what if artificial intelligence could perform this translation automatically?

This is where **LLM4OR** comes into play—a Deix research project exploring the use of **Large Language Models (LLMs)** to interpret problems described in natural language and convert them into **ready-to-use mathematical models**.


*The idea is as simple as it is ambitious: Can we teach AI to perform our analysis and formalization work?*

Not to replace human intelligence, but to make it more efficient and faster.

### The Core of the Project: Language Models and Optimization

**How Does an LLM Work?**

A **Large Language Model (LLM)** like ChatGPT does not "think" or "understand" in the human sense of the terms. It **does not reason like an optimization expert** who analyzes a problem, reflects on its implications, and consciously formulates a mathematical model.

Instead, an LLM functions as a highly advanced sentence completer. It works by associating word sequences with other statistically probable sequences based on all the training data it has seen.

*A useful analogy?*

*Imagine a tourist in a foreign country who doesn’t speak the local language. They have a phrasebook with pre-translated sentences: they don’t truly understand the meanings but know that to ask for the check at a restaurant, they must say, "The check, please." If the situation changes slightly (e.g., a waiter asks if they want to pay by card or cash), they might be confused.*

An LLM works in a similar way: it doesn’t truly "understand" the problem, but it knows which type of response is most probable based on its training data.

**LLMs for Operations Research**

Despite this statistical approach, LLMs can be surprisingly effective. To test whether an LLM can simulate the ability to translate an optimization problem into a mathematical model, we:

- **selected an open-source model** (Llama3-8B-Instruct) to evaluate its effectiveness compared to proprietary models like ChatGPT,
- **used a public dataset** containing optimization problems already translated into mathematical models, providing the model with clear examples,
- **created a custom dataset** with more realistic and complex problems to further improve the model's capabilities,
- **fine-tuned the model**, allowing it to "study" the new data to enhance its translation accuracy.

It's important to clarify: we did not "teach" the model to think like an expert but trained it to recognize and replicate effective translation patterns.

A system capable of converting problems described in natural language into mathematical models.

### The Results: AI vs. Human Experts?

We compared our model with ChatGPT-3.5 and ChatGPT-4o on a set of problems, evaluating:
- correct identification of variables and the objective function
- precise translation of mathematical constraints
- ability to generate executable code for solving the problem.

Our model outperformed ChatGPT-3.5 and, thanks to fine-tuning on custom data, even **surpassed ChatGPT-4o's capabilities**.


*But what does this mean in practice?*

*It means that a user could describe a problem in plain language and automatically receive a model ready to be solved by optimization software.*


### Applications: Where Can This Technology Be Used?

Today, the project is still in an experimental phase, but its applications are vast:
- **Supporting our experts**: At Deix, this technology could **accelerate our work **by reducing the time required for the initial modeling phase.
- **Integration into [JINN](https://deix.ai/solutions/products/jinn)**: Deix's AI platform for creating multi-purpose AI agents now includes features for training and fine-tuning a dedicated model based on an open-source one, not just for optimization problems.


### A First Step Toward Optimization Automation

The LLM4OR experiment has demonstrated that LLMs can support optimization by translating problems into mathematical models with increasing accuracy.


🚀 This isn’t about replacing human intelligence but enhancing it, providing a tool that speeds up and simplifies the work of experts.

💡 For Deix, this reaffirms our mission: exploring AI's potential to create innovative and practical solutions.



A Deix project exploring the potential of language models in optimization

LLM4OR: The Generative AI That Translates Problems into Solutions

Two days of innovation, connections, and new opportunities: our team participated in the Innovare X Crescere event, organized by ELIS Innovation Hub in Rome, and we returned with fresh ideas, new collaborations, and—above all—a great sense of pride in the path we are building.

**Innovation Comes Through Collaboration**

Innovare X Crescere is an event that connects corporations, startups, and innovative players, creating opportunities for discussion and collaboration to drive forward concrete projects. Within this framework, Deix had the opportunity to present the work carried out as part of the Open Italy program: [a Proof of Concept with Gruppo Lactalis Italia to test how generative AI can transform the world of industrial maintenance in real-world scenarios](https://www.linkedin.com/posts/gruppolactalisitalia_impact-award-activity-7303733768064794625-i8Eo?utm_source=share&utm_medium=member_desktop&rcm=ACoAACXkgLsBwaBxzk15U3w6dDrLBauc2YhJW-s).

**The Project**

Our goal? To demonstrate that generative AI can support field operators, making maintenance activities more efficient and intuitive. Thanks to our solutions, we aim to:
- Facilitate and speed up maintenance reporting by allowing technicians to record a voice note.
- Simplify knowledge management by structuring data and procedures in an easily accessible way.
- Provide real-time information to operators, reducing search times and improving response speed.

**An Ecosystem That Fosters Innovation**

A big thank you to ELIS and the Open Italy program, which continues to be a powerful innovation accelerator, connecting startups and large companies to bring high-impact projects to life.


We are only at the beginning of a journey that will redefine the future of smart maintenance. We can't wait to share the next steps with you!

Generative AI and Smart Maintenance for the future of industry

Deix at Innovare X Crescere

In a world where decisions are often made in a rush and words dominate our way of communicating, stopping to think with our hands might seem like an unusual idea. Yet, that’s exactly what we did this morning at Deix, guided by the amazing Elisabetta Maier, through a session of LEGO® SERIOUS PLAY®.

It was an experience that led us to reflect on our roles within the team, what success means to us, and the future Deix we want to build. And we did it not through presentations or long discussions but with LEGO bricks, transforming thoughts and visions into tangible models.

**Thinking with Your Hands: The LEGO® SERIOUS PLAY® Method**

The LEGO® SERIOUS PLAY® method is a facilitation technique that helps individuals and teams explore complex ideas through physical construction. Based on the concept of "building to understand," this approach allows us to visualize abstract concepts, stimulating creativity, collaboration, and strategic thinking.

The method follows four key steps:
- *Challenge* → The facilitator presents a question or problem.
- *Building* → Each participant responds by creating a model with LEGO bricks.
- *Sharing* → Everyone explains their model and the meaning behind it.
- *Reflection* → The group discusses and builds a shared vision.

Beyond its apparent simplicity, this process leads to deep insights and remarkable team alignment.

**What Did This Experience Teach Us?**

During the session, we found ourselves shaping our thoughts, discovering new perspectives, and building visual metaphors to express who we are and where we want to go. Here are some key takeaways:

* The power of metaphors – A LEGO model can tell a story more effectively than a thousand words.
* Everyone has a voice – The method puts everyone on equal footing, fostering open dialogue.
* Building helps us see – Translating an idea into a tangible object clarifies and strengthens its message.
* The future is built together – Envisioning the Deix of tomorrow requires collaboration, vision, and creativity.

**What’s Next? Let’s Keep Building!**

This experience showed us the importance of pausing to reflect, exploring ideas in new ways, and building a shared vision together. And we can’t wait to put what we’ve learned into practice!

Our journey with LEGO® SERIOUS PLAY®

Let Your Hands Think

2024 has been a year of milestones, challenges, and growth. Yesterday, we had the opportunity to celebrate our successes during the end-of-year plenary meeting. This special moment allowed us to reflect on what we’ve accomplished together and to recognize the commitment and dedication of our entire team, who contribute every day to building something unique and innovative.

Among the most significant moments of 2024, in addition to the many successful projects carried out with our clients, we celebrated the awards received for our solutions (such as the [DigiSafeWell Award](https://www.deix.ai/en/magazine/articles/deix-wins-the-digisafewell-award) and the [InfoCamere challenge](https://www.deix.ai/en/magazine/articles/winners-of-the-infocamere-data-driven-innovation-challenge)), the launch of new partnerships, and [industrial research projects](https://www.deix.ai/en/magazine/articles/deix-and-egeos-in-collaboration-with-esa) focused on developing cutting-edge solutions.

A heartfelt thank you also goes to our partner OTTOZEROUNO, who supported us with great professionalism and collaboration, contributing to the success of this intense and rewarding year.


An extraordinary year with the Deix team

Deix Plenary 2024

**The Importance of Time Series**

Time series lie at the heart of many critical real-world applications. Financial data, energy consumption, user behavior data, industrial measurements, and weather forecasts—all of these scenarios generate information structured over time. Time series, in essence, represent data sequences recorded at regular intervals, forming one of the most pervasive and valuable data formats in modern applications. Their analysis enables understanding phenomena, identifying recurring patterns, and making predictions—essential for strategic decision-making in numerous sectors.

For businesses, understanding how a system will evolve over time or how a specific variable will influence future behavior can mean the difference between success and failure. This is why developing models capable of managing and interpreting these dynamic data effectively is crucial. Here, the power of advanced AI models like Transformers comes into play.

**The Transformer Revolution**

Transformers are a special type of neural network designed for processing data sequences. They have emerged as one of the most groundbreaking innovations in artificial intelligence in recent years, particularly in natural language processing (NLP). First introduced in 2017, these models have revolutionized how AI handles NLP, giving rise to advanced systems such as large language models (LLMs)—including GPT, BERT, and many others.

The core idea behind Transformers is the "attention" mechanism, which allows the model to focus on different parts of an input flexibly and effectively. By simultaneously considering relationships across all positions in a sequence, Transformers have proven extremely powerful in capturing long-term dependencies in sequential data. This makes them particularly well-suited not only for text but also for time series analysis.

**Why Transformers for Time Series?**

Time series can be viewed as sequences of data that evolve over time, much like text is composed of a sequence of words. Traditionally, models like recurrent neural networks (RNNs) and Long Short-Term Memory networks (LSTMs) have been used to model time series due to their ability to maintain temporal memory. However, these approaches face limitations when dealing with particularly long sequences or complex dependencies.

Transformers, with their self-attention mechanism, overcome these limitations by providing a broader and more contextualized view of the data. With Transformers, every point in a time series can relate to every other point, without the need to process the sequence rigidly as in RNNs. This enables Transformers to automatically and quickly identify complex patterns, even across significant time spans.

**Embedding and Prediction: How Transformers Work for Time Series**

Applying Transformers to time series involves several key steps. First, the time series are represented through embeddings—dense numerical representations capturing the salient features of the sequences. These embeddings serve as inputs to the Transformer model, which, through its attention layers, identifies patterns, trends, and relationships within the data.

This highly informative representation can be utilized in classification or prediction layers, depending on the final objective. If the goal is to classify an event (e.g., detecting anomalies in industrial data), the embeddings are fed into a classification layer. If the aim is to predict future values (e.g., monthly sales forecasts), the embeddings are used in a prediction layer.

Thanks to these characteristics, Transformers can tackle a wide range of time series analysis problems, enhancing both the accuracy of predictions and the ability to interpret underlying phenomena. Additionally, Transformer-based architectures can integrate external information or diverse contexts, further boosting the models' effectiveness.

**Opportunities for Businesses**

Using Transformers for time series analysis marks a significant step forward compared to traditional methods. With the ability to capture complex patterns and long-term relationships, these models offer new possibilities for forecasting and classifying temporal data, making AI applications in business and industrial contexts more effective.

At Deix, we have experience applying this approach to financial data classification to detect potential fraud and to forecasting time series from IoT sensors to estimate the remaining useful life of industrial components. However, the potential applications are even broader.

In a data-driven world, fully understanding and leveraging the power of time series provides businesses with concrete tools to improve their competitiveness and decision-making capabilities. Transformers stand out as one of the most promising solutions to address these challenges innovatively and effectively.

Using Transformers for Time Series Analysis

On November 28, 2024, we had the honor of participating in Matcher On Stage, the concluding event of Matcher Data-Driven Innovation - 2024 Edition. Organized by Art-ER and the Enterprise Europe Network, this international open innovation program aims to foster strategic business meetings between top companies from the Emilia-Romagna Region and international startups or scaleups.

Our Jacopo Corno, Head of Business Solutions at Deix, was among the key speakers at the roundtable titled "Data Analysis as a Key to Strategic Decisions." The event featured prominent organizations such as Fondazione ITL, Chiesi, Gruppo CAMST, and Rekeep, which together discussed the crucial role of data analysis in transforming business decisions into strategic opportunities.

The panel explored real-world cases and innovative solutions showcasing how leveraging data can enhance companies' decision-making capabilities, boost their competitiveness, and promote future-oriented strategies.

Deix's participation in Matcher On Stage marks another step forward in our commitment to driving innovation through Mathematical Intelligence and collaboration with prominent partners.

Deix at Matcher 2024

We are pleased to announce that Deix Srl has secured funding through a cascading grant from the PNRR, under Spoke 2 "Integrative AI" of the FAIR (Future Artificial Intelligence Research) project. This initiative, coordinated by the Bruno Kessler Foundation, aims to foster the integration of heterogeneous techniques that leverage both data-driven and model-based approaches, developing theories and solutions capable of addressing the challenges of Artificial Intelligence. Integrating mathematical models with data-driven techniques represents the key challenge of this call, a vision that reflects Deix's distinctive approach to Mathematical Intelligence: a blend that combines the rigor of mathematical modeling with the flexibility and power of data-driven AI.

Deix's project, named INTEGRA (INtegrative AI Technologies for Ergonomic Guidance and Risk Assessment), aims to improve workplace safety by combining advanced Artificial Intelligence techniques with physical models of human motion. This synergy will enable the automation of ergonomic risk assessments, using standard scales such as REBA, RULA, and OWAS, leveraging data acquired through motion capture technologies and image-based pose estimation.

Collaboration with the Bruno Kessler Foundation, a leader in the "Integrative AI" Spoke of the FAIR partnership, will provide Deix with the opportunity to develop cutting-edge solutions, benefiting from the expertise and resources of a center of excellence in Artificial Intelligence.

INTEGRA is a natural evolution of our [Varda-AI](https://www.deix.ai/en/solutions/products/varda-ai) platform, already used to monitor behaviors and movements in industrial environments. The integration of new functionalities for real-time ergonomic evaluation will reduce workplace injuries and improve workers' health, while also offering solutions applicable to fields such as healthcare and sports.

This project strengthens Deix's commitment to combining Artificial Intelligence and mathematical modeling to develop innovative and sustainable solutions, contributing to the digital transformation of companies and promoting a culture of prevention and workplace safety.

Innovation in Ergonomics Powered by Artificial Intelligence

Deix Launches the INTEGRA Project

On October 29 and 30, Deix participated in SMAU Milan, a key event for innovation in Italy, hosted at FieraMilanoCity. With the support of the Lombardy Region, we were present with our booth alongside other local excellence, representing the best of Lombard innovation.

The experience at SMAU allowed us to meet with prominent corporations interested in our Mathematical Intelligence solutions and our expertise in AI and mathematical modeling. These meetings opened up collaboration opportunities with organizations that share our goal: accelerating companies' digital transformation in an ethical, sustainable, and transparent way.

Participating in the workshops dedicated to innovation and artificial intelligence was fundamental for us to engage with emerging trends and, above all, with the main pain points of companies. Among the critical challenges discussed, the adoption of AI tools within companies stood out: bringing innovation also means overcoming cultural and structural barriers to ensure the effective and advantageous implementation of advanced technologies.

We also had the opportunity to participate in two speed-pitching sessions, where we presented our solutions to an audience of industry experts and decision-makers.


The two days at SMAU were intense but extremely productive, thanks to the entire Deix team—Jacopo, Fabio, Rossella, and Francesca—and our partners at Ottozerouno, Luca and Marco, for the great work done!

Articles and news