Artificial Intelligence
AI in the work, not in the pitch.
When a client asks me to add AI to their application, the first question isn't 'which API?' — it's 'does this actually solve the problem?' I've been integrating LLMs and agents into web applications since these tools became operational, and I use AI daily in my own development workflow, on demanding industrial projects.
Areas of work
AI-augmented development
I use GitHub Copilot, configured agents (rules, skills, hooks), and automated workflows in my daily delivery cycle. Across ~10 applications developed continuously for an industrial client since 2021, AI is embedded in the development chain — review, generation, validation. This isn't an experiment: it's the working mode.
LLM/AI integration in web applications
I integrate LLMs via API (OpenAI, Anthropic, Mistral) into existing applications: AI-assisted simulation, decision support, tutors, contextual assistants. For e-xode Learning, an online learning product I built and shipped to production, I chose a lightweight API integration over a dedicated model — the volume didn't justify fine-tuning costs, and inference latency was acceptable through the standard API.
AI advisory and arbitrage
Before plugging in an API, I assess whether AI is the right lever. If it is: RAG or fine-tuning? Lightweight integration or dedicated model? Which API, for which use case, at what operational cost? This arbitrage work belongs to the architecture mandate — it shouldn't be delegated to an external AI consultant who doesn't know your codebase.
My approach
I distinguish three situations: AI delivers real value, AI is feasible but not the priority, and AI is the wrong answer. The third is the most common — and recognising it is the only way to avoid a wasted investment.
When AI genuinely fits
AI is relevant when it reduces a real friction: a user searching for an answer across 500 pages of documentation, a business rule too complex for a classic rules engine, a signal in unstructured data. In these cases, an API integration with solid prompt engineering delivers measurable results within a few weeks.
When I say no
If your problem is solvable with a well-crafted SQL query or a properly designed filter, adding an LLM will only introduce latency, cost, and an extra failure point. I've declined LLM integrations on projects where the real fix was a better data architecture. That's not a stance — it's what I would do for my own application too.
Concrete experience
Two cases I can describe — one at a general level for confidentiality reasons, the other in detail because it's my own product.
Industrial client — ~10 applications (since 2021)
Saint-Gobain (Glass + Gypse): AI is integrated into my development workflow across all applications in the engagement — Copilot, agents, validation hooks. One application for industrial plate dryer simulation includes an AI layer for operational guidance. Business details are covered by confidentiality, but the usage is daily and operational.
e-xode Learning — AI tutor (live)
I built and launched an online learning product: e-xode Learning. The AI tutor is present at every exercise — it answers the learner's questions from the course reference material, in multiple languages, with no usage cap. I chose an integration via the Anthropic API with structured context rather than a fine-tuned model — the content base evolves too fast to justify regular fine-tuning.
Deadweight — configuration audit (open source)
The governance I apply to my own AI-assisted delivery, published and open to inspection: the same auditor that checks every CLAUDE.md, skill and rule I write, released under MIT instead of kept to myself. It runs on this site's own configuration - 0 errors and 1 warning at the last floor reset. The code reads, runs, and checks out.
A project with an AI component?
Describe the problem you want to solve — not the solution you have in mind. I'll tell you honestly whether AI is the right tool, and if so, which approach.
Let's talk about your project