Background
My name is Ilan Da Silva. I build artificial intelligence systems from Liège. I started early and through the technical side: during my studies, I integrated AI into the management software of a Belgian publisher, automating the encoding of catalogues with several thousand references.
In parallel, I trained my own code language model from scratch, SmallCoder, with the support of Google's TPU Research Cloud program: a 303-million-parameter model that approaches models twenty times larger on code generation. Today, I design and deploy internal AI systems in production for companies and regulated professions.
My approach
I don't ship a chat in a window. I build systems that live inside your real work, connected to your tools and your files. Three principles guide everything I do: privacy (your data stays with you, hosted in Europe), traceability (every output is verifiable) and human review (you keep control over the decisions that matter). Because I've trained a model from scratch, I know what an AI system can really do, and, above all, what you shouldn't trust it with blindly. That's the difference between integrating AI and simply opening a generic tool next to Word.
Case study 1, regulated consulting firm (anonymized)
- Problem: the firm was already using a generic AI tool daily, but with no integration to its files. Every task meant reconfiguring everything, quality depended on the person, and confidentiality was never guaranteed.
- System delivered: an internal, agentic AI platform organized around the case file. An orchestrator distributes work across specialized agents (drafting, deadlines, engagement letters, client onboarding, billing), connected to their document environment, on a closed network hosted in Europe, with human validation at every step.
- Impact: drafting went from several days to an immediate structured first draft, with consistent quality regardless of the person, and data that never leaves the firm's environment.
Case study 2, SmallCoder (language model)
- Problem: it's assumed that a small model can't compete with large ones on code generation.
- System delivered: training a 303-million-parameter model dedicated to code from scratch, with the support of Google's TPU Research Cloud program.
- Impact: state-of-the-art code performance for its class, close to Mistral 7B on the HumanEval benchmark while being roughly twenty-three times smaller, and a model picked up and reused by the community.