What we learn from building data and AI systems, published as we go.
Private inference
In regulated environments the bottleneck is no longer raw model intelligence. It is where inference runs, who can see the data and what each task costs.
Didier Franc, September 2026
Independence
Committing to one vendor’s ecosystem limits what you can build, what you pay and how well you recover. Neutrality is the foundation for the opposite.
Didier Franc, January 2026
Agents
A controlled study found that a single coding agent with a minimal harness matched or beat elaborate multi-agent frameworks. The best architecture may be the one that gets out of the model’s way.
Didier Franc, November 2026
Inference
DSpine drafts several tokens at once that actually belong together. Integrated into Lisa, it speeds up generation by 2.7 to 3.8 times.
Didier Franc, November 2026
