Engineering notes
Practical notes for deciding what to build, what to audit, and what to avoid before an AI project reaches production.
Repair an AI-generated app: scope the first fix
How to scope a first paid repair for an app built with AI, reproduce the failure and decide whether a rewrite is actually needed.
Subcontracting an AI integration: brief, acceptance and handover
How an agency can assign one AI module with clear ownership, acceptance checks and a handover its team can use.
Retry agent tool actions without duplicating their effects
Use persistent operation identifiers, explicit unknown outcomes and reconciliation to control retries of external actions in LLM integrations.
Prevent cross-tenant document leaks in RAG
Check access before retrieval and test document leakage through cached answers, citations and source links in a multi-tenant RAG application.
AI consulting in Madrid: choose the work you need
How to choose an AI review, a scoped integration or technical leadership, with deliverables you can assess before agreeing a project.
What a RAG and agent audit should deliver
Scope an AI audit around a release decision, with reproducible findings, permission tests and concrete acceptance criteria for remediation.
Remote AI consulting: agree the work, reviews and handover
A working agreement for remote AI projects across Europe and Latin America: ownership, time zones, review material, access and technical handover.
Starting a remote fractional AI CTO engagement
Define a first fractional AI CTO engagement through decision ownership, technical reviews, hiring responsibilities and a clear point to reassess the work.
One production agent: tools, state and execution limits
Design an agent runtime that authorises tool calls, records operation outcomes and recovers from failures before allowing external changes.
Fractional CTO, adviser or engineer: which role do you need?
Choose between implementation, technical advice, fractional leadership and a full-time CTO by examining the work and decisions your startup needs covered.
RAG or fine-tuning: diagnose the failure first
Separate missing evidence from inconsistent model behaviour, then compare retrieval, prompting and fine-tuning against a defined acceptance test.
When dividing work between agents earns its cost
Assess parallel work, separate permissions and agent handoff contracts against their coordination cost before adopting a multi-agent architecture.