Reducing the cost of an AI application: model routing, caching, context, and FinOps
The best optimization is not about choosing the cheapest model. It is about avoiding unnecessary calls and using the right capacity for each step.
PartITech blog
Technology blog
Technical analysis, practical guides and field experience to build, modernize and maintain sustainable applications.
Prompt injection and AI agent security: threats, attack scenarios, and defense in depth
No system prompt can guarantee that a model will always ignore a malicious instruction. The defense must prevent a misinterpretation from becoming a leak or an action.
SEO, GEO, and AI answer engines: making expertise visible without producing artificial content
GEO does not erase SEO. Answer engines still need pages that are accessible, understandable, reliable, and original enough to deserve being cited.
Long-duration agents in sandbox: architecture, security, and task recovery in 2026
An agent who works for hours needs checkpoints, budgets, persistent states, and access limits. The model is only part of the system: the harness and the sandbox determine what it can actually do.
Real-time voice agents in 2026: architecture, latency, consent, and quality of service
The voice does not forgive dead times, ambiguities, or poorly confirmed actions. A voice agent is both a real-time, conversational, and business system.
Big bang overhaul or gradual migration: how to choose without putting the business at risk?
The risk of a revamp is less concentrated in new code than in the transition: data, users, integrations, coexistence, and rollback.
API-first and composable architecture: gaining scalability without building an unmanageable system
Decoupling only creates value if the boundaries, contracts, and responsibilities are more stable than the system they replace.
Core Web Vitals and web performance: building a measurable budget rather than chasing a score
Performance is not a one-time cleanup. It is a design constraint measured on the devices, networks, and paths actually used.
Open AI models, self-hosting, and edge AI: what is really possible in 2026
Downloadable weights provide more control, but not automatically a sovereign, economical, or maintainable solution. The architecture must be tested on the real task and hardware.
Sovereign Generative AI: API, Private Cloud, or On-Premise — Which Architecture to Choose?
Sovereignty is not limited to the location of a server. It combines data control, dependencies, contracts, models, skills, and exit capacity.
Designing a secure and useful client extranet: architecture, rights, documents, and user experience
An extranet is not a site to which one adds a connection. It becomes a boundary of the information system and must be designed as a business product.
Updating or migrating an application without interruption: compatibility, data, and rollback strategy
Zero downtime is not a deployment command. It is a compatibility property between versions, data, users, and external systems.