Research & experimentation

Research & experimentation

Alongside client deliverables, since 2022 I have been conducting independent research into neural architectures and multi-agent orchestration systems. This work is not yet published — I prefer to validate rigorously before sharing.

This page exists to give an anonymized overview of the topics explored, without revealing architectural details. They are mentioned here to provide context for clients wondering “where does this AI applied expertise come from?” — the answer: “from years of building my own systems from scratch, not just calling APIs.”

Active domains

Custom neural architectures

Designing alternatives to standard Transformers: hybrid combinations of liquid time-constant networks (LTC), learnable splines (KAN — Kolmogorov-Arnold Networks), and modified attention mechanisms. PyTorch from-scratch implementation, validation against public benchmarks, inspired by recent papers (NeurIPS, ICLR).

Why? Because standard Transformers reach their limits in some data regimes. Understanding alternatives = understanding when a client truly needs a fine-tuned generic model, vs designing something more targeted.

Multi-agent orchestration and delegation

Studying protocols that let multiple specialized models collaborate on a complex task — a structured approach rather than one large model doing everything. Typical workflow: a planner model, executor models, a verifier model. The context pollution anti-pattern is studied in practice.

Why? Because in 2026 most production AI problems are no longer “which model?” but “how to orchestrate multiple models?”. This skill transfers directly into RAG or agent deliverables for clients.

Peer-to-peer distributed compute

Research into protocols allowing multiple machines (laptop + cloud GPU + volunteer machines) to collaborate on training or inference without uploading model weights. Inspired by DiLoCo, federated learning, and modern hub-and-spoke architectures.

Why? Because the cloud GPU bill has become the limiting factor of independent R&D. Learning to train smartly with less, rather than renting more.

Automatic LLM output verification

Designing pipelines where an LLM output is verified by a second LLM (or a formal system) before being presented to the user. Includes: hallucination detection, citation verification, comparison against ground-truth.

Why? Because shipping an AI system to production without verification = shipping a risk. This discipline directly feeds my Loi 25 audits (where no silent error is acceptable).

Publication status

Work not publicly published. Available for discussion under NDA with serious clients who want to understand the depth of the expertise before a mandate.

No self-promotion on arXiv or public GitHub — neural architecture research is highly competitive, and several ideas still need to be validated/protected.

Going further

If you represent an organization that could benefit from these skills (applied R&D, third-party AI system audit, custom design), contact me directly. An initial conversation is always free and can quickly clarify whether a mandate is relevant.