Recognized for its research projects in artificial intelligence, ILLUIN Technology has developed support services for large groups, but also a suite of products. Starting with Dialogue, a tool developed from 2018 to design conversational agents. It has been deployed on a large scale by large French groups, including Enedis, EDF, La Macif, Boursobank, Randstad and Geopost. The company has around sixty customers.
Since then, ILLUIN has expanded its offering with various services dedicated to intelligent document processing – Doc Automation, Mail Automation, Search – but also voice analysis and hot and cold customer intent – Voice Parser and Speech Analyzer .
“In the intelligent document processing segment, our major clients have submitted us to fairly solid benchmarks,” says Robert Vesoul, co-founder and CEO of ILLUIN Technology. “Our solutions can handle the most complex documents: a hospital bill, a pay slip, a car accident report, etc. “.
The names of these products have evolved. They themselves have largely been “remodeled” under the influence of generative AI, according to the manager.
However, there has been one constant for four years. They all rely on a main “backbone”. “We decided in 2023 to make it a full-fledged product,” says Robert Vesoul. “This is the framework that we open to all of our clients to use in cases of development, tailor-made production, AI and Generative AI projects.” This framework is the nAIxt platform. It is presented as a “low-code studio” intended to cover the design of applications involving the orchestration of AI models, generative or not. The scientific community speaks of a composite AI system.
While Copilot and ChatGPT have dominated the scene for the past two years, some companies are hoping to benefit from the additional automation LLMs bring to business processes. “I was recently in a talk at the Positive AI summit. There were speakers who mentioned the fact that the most advanced organizations in generative AI are those that truly develop tailor-made applications, their core business,” says Robert Vesoul.
A lack of specific tools
However, the tools available to do this remain complex or limited. Frameworks like LangChain and LlamaIndex have established themselves, but the landscape of tools is growing and these libraries require the intervention of developers or people familiar with Python or JavaScript. “These open source solutions evolve quickly, but cover part of the subject, and their development has been uninspired with regard to the integration of corporate information systems,” believes Robert Vesoul. LangChain also exists in a more complete commercial distribution in this area than its open source project.
Historically, the majority of analytics and data science teams have members who are not trained in Python and JavaScript. These and externals use or have long used tools with a WYSIWYG interface to manage data pipelines. Treating data like code remains a relatively new concept and the barrier to adopting DevOps/MLOps practices remains high.
“Orchestration of a framework in Python in an organization that brings together data scientists, data engineers and data architects constitutes a major challenge,” illustrates Robert Vesoul. “Few companies manage to do this effectively.”
Cloud providers are trying to remedy this, but certain companies, including some of ILLUIN’s customers, in addition to proprietary lock-in, fear integration conflicts between sometimes competing solutions. “Some of these tools in the different clouds depend on biases, which are not necessarily those of the organization for other development projects,” notes Robert Vesoul.
Added to this is the need to manage several major language models, but also different algorithms of all kinds. “There is the question of performance, but also of costs,” he adds. “Today, companies in 2024-2025 have listed (or are finishing listing) priority use cases. They must now organize, develop, orchestrate this across the organization.” And the first box to check is none other than that of token consumption: the performance-cost ratio.
“Most of the large clients we question tell us, “one of our challenges is to take advantage of these models, but while remaining independent”,” relates the CEO of ILLUIN Technology.
Agnosticity, the key word
This is one of the startup’s key arguments to convince large groups. Its platform must orchestrate workflows powered by “hybrid” AI in an agnostic manner, whether on the three major cloud infrastructures (GCP, Azure, AWS), on sovereign and qualified SecNumCloud clouds (including OVHcloud), but also on site. “This is about proving that we can operate use cases without the use of OpenAI APIs, quickly deploy capabilities and also change technological components from different suppliers during the course of a project,” specifies Ghislain Jeanneau, nAIxt product director at ILLUIN Technology.
Technically, ILLUIN customers and partners meet there. This is the technical base of the products mentioned above. It’s also a familiar solution for Data & AI teams.
According to Ghislain Jeanneau’s presentation, nAIxt structures a project in two phases. First there is configuration in the studio, then execution, that is to say the transformation of data “into usable results” via deployed APIs.
Users manipulate projects including reusable objects – prompts, vocabulary, etc. “Project templates allow you to quickly start working on frequent use cases, such as RAG, document processing, an agent-oriented paradigm,” etc. », assures Ghislain Jeanneau.
Development in nAIxt revolves around two types of objects. Firstly, there is a pipeline type object allowing you to describe drag-and-drop processing in just a few steps. The “test sets” object brings together test data on which the pipelines will be tested with each modification. “With each change made, the pipeline is executed again and allows us to understand the impact of the changes made.”
Once the pipeline is ready, the “deploy” button exposes this pipeline via an API. An interface and modules allow performance and explainability of results to be evaluated in the configuration phase and during execution. The UI includes a cost estimation function to determine through iterative phases of pipeline testing which are the most relevant AI models to leverage. In the event of an error, problems or for CISO purposes, the platform keeps audit logs.
On paper, this all looks like an ETL/ELT (data extraction, loading and transformation) tool. But here, the T would not only correspond to the query engine of a database or an Apache Spark type technology, but to different transformation modalities, embodied by the inference of AI models (LLM, VLM, computer vision, NLP, etc.). These transformations can take place at different places in a pipeline to structure the results and can be linked together. They are performed by operators, specific functions that call different LLMs and tools. ILLUIN says it has more than 150 operators. The code generated by the platform can be executed through functions in “serverless” mode on third-party FaaS services.
If it eventually involves creating a runtime, companies can access the source code of the pipelines as needed. The visual programming tool arbitrarily generates TypeScript code. “We chose this language because it can handle highly asynchronous workloads,” explains the product manager.
nAIxt integrates with various platforms including Amazon Bedrock, Google Vertex AI, Azure AI, Microsoft Sharepoint, Google Drive, Salesforce, Dynamics 365, vector databases (Qdrant, Elasticsearch), model inference tools (vLLM , TGI, Ollama), etc.
The documents themselves can be manipulated in the studio and an explainability system aims to link the information to its sources. Some operators applying the K nearest neighbors method make it possible, for example, to verify that an amount displayed in an invoice is close to that announced in a contract. Here, ILLUIN recommends using vision language models, a RAG architecture and its ColPali technique, exploiting the Google PaliGemma “openweight” models to create visual embeddings allowing more faithful extraction of data from PDF or JPG .
“Generative AI use cases often focus on conversational applications, where the end result is generally a simple character string or JSON,” notes Ghislain Jeanneau. “Tools allowing advanced pre-processing or post-processing of data according to specific conditions are lacking,” he assures. “This limits the possibilities to end-to-end automations, often geared towards specific workflows. With generative AI, these scenarios are less direct than traditional approaches, making orchestration tools essential to achieve a higher level of automation.”
ILLUIN Technology believes it has found a way to stand out
French players or those of French origin say they are ticking the agnostic cache, notably LightOn and Dataiku. “My understanding is that Lighton is more focused on wizards, user interaction and simple task chaining, while Dataiku, from a more traditional DNA, remains focused on model training, MLOps and classic analysis” compares Robert Vesoul. “NAIxt seeks to unify these approaches with a development studio capable of handling the full range of generative AI use cases, from back-office pipelines to conversational tools, where existing solutions are often limited.” The ILLUIN platform would then be complementary to these two solutions. And rather ahead of the market.
Still, players like Microsoft, AWS, Google Cloud and Salesforce want or can handle processes of this type. For now, they have further strengthened the assistants with RAG tools and architecture to make them a sort of “ShivaGPT”. During the official presentation of nAIxt on December 4, Geopost presented its intention to use generative AI to extract and analyze parcel-related events from IS logs, in order to produce actionable summaries. These summaries can be presented to advisors to respond to customers or directly integrated into interactions via conversational bots, without necessarily using a RAG architecture. “Organizations that are mature on these subjects are now looking for suitable tools to implement this type of end-to-end process,” insists the CEO.
For now, ILLUIN continues to target “large public and private organizations”, but hopes to bring together a broader community by offering an open source distribution of nAIxt in 2025. Also a way to expand into the ETI market.
