Business Method
The shiny toy
The AI project that dazzles in the demo and dies in a drawer has a name and an antidote. A three-signal test to tell a tool from a whim.
Read more
Blog
What we discover while building AI in production with our clients.
Security Architecture Agents
Four layers between your company and the one next door
One assistant answers several companies and none of them can see another. The four layers that guarantee it, a lesson about permissions and the redesign that erased a whole class of failures.
Read more
Agents Architecture Security
Here is a secret: I don't like agentic architectures
We sell AI agents, and we don't like how almost everyone builds them. Business rules cannot live in the prompt, they have to live in the code.
Read more
Data Architecture Agents
Data before prompts
An entire market sells prompt engineering. Our production experience says you gain more by organizing the data than by polishing the prompt.
Read more
Observability Data Agents
The silent zero
A zero is not a neutral answer, it is an unclassified alarm. The three causes of a "no data" response and why the third one is the most treacherous.
Read more
Maintenance Observability Business
Starting an AI project is easy. Keeping it alive, nearly impossible
Brilliant AI projects launch well and die young. Profitable ones are easy to start and easy to maintain. The difference is decided before any code is written.
Read more
Observability Maintenance Agents
The symptom signature beats the fix
Documenting how a failure is recognized from the outside pays more than documenting how it was fixed. Three real signatures from an AI assistant in production.
Read more
Agents Data Observability
The auditor that trusts no figure
A second model that distrusts by design, a deliberately short memory, and the metric that catches an assistant answering from memory.
Read more
Agents Reliability Maintenance
What your assistant does when a tool goes down
A conversational assistant depends on systems that fail. The circuit breaker that protects the user, and the three lessons production taught us about it.
Read more
Business Method Automation
Measure AI by profit, not by productivity
Productivity is AI's comfortable metric, it always looks good and commits to nothing. We prefer the uncomfortable one, which concrete gain did the system move.
Read more
Agents Knowledge Product
Two kinds of "no" that are not the same
When an AI assistant cannot answer, there are two causes of a different nature. Confusing them is the classic mistake that inflates documentation without fixing anything.
Read more
Agents Quality Observability
Detecting and preventing AI hallucinations is architecture
A hallucination is a well-written, false answer. The four mechanisms our systems use to corner them in production, with their numbers and their scars.
Read more
Data Agents Security
The agent that queries your data does not write SQL
The text-to-SQL pattern promises a model that writes queries. We built it for Savian with the opposite decision, and that decision is what makes it safe.
Read more
Automation OCR Back office
Invoice digitisation with OCR and AI, from a case in production
What changes when OCR meets a language model, how it looks in a real case with utility invoices and why validation is the actual product.
Read more
Agents Real estate Automation
An AI agent for real-estate agencies, from the inside
Dozens of WhatsApp messages a day, five to ten minutes per inquiry and an overloaded team. What the agent that filters requests for a Barcelona agency actually does.
Read more