Local AI and open-weight models: does data really stay private?
Running a model locally reduces some transfers, but it does not automatically make data, RAG and logs private. A guide to choosing cloud, on-premises or hybrid.
Read the article
INSIGHTS · ZENCODE
Practical guidance for navigating processes, software and AI.
Guides and insights into automation, integrations and development: what to consider before choosing a tool or starting a project.
Articles in the Artificial intelligence category
MCP and permissions: connecting AI to data without unnecessary access.
A secure MCP connection does not depend on a promise in the prompt: tools, credentials, data, network controls and approvals must enforce verifiable boundaries.
Read the articleVibe coding: from prompt to production without losing control.
A prompt accelerates the first prototype; requirements, operational boundaries, independent checks and rollback make it manageable software.
Read the articleThe test passes, the maintainer rejects it: the limits of coding agents.
Automated tests are necessary, but they cannot measure a patch’s full correctness, architectural consistency and future cost on their own.
Read the articleWhen a coding agent deletes data from a database: rollback comes before the incident.
A database is not a file to regenerate: isolation, cautious migrations, backups and verified restoration must come before agent autonomy.
Read the articleAround one million lines with Codex: why the method matters more than the prompt.
The OpenAI case does not celebrate a perfect prompt: versioned documentation, verifiable architecture and feedback loops made Codex effective.
Read the articleAI in software: from demo to dependable process.
Evaluation, data, prompt injection, privileges and oversight: how to take AI beyond a convincing demo.
Read the article