AI agents, automation and judgment inside Autodesk products: Vault, Inventor, AutoCAD and Forma. What I build and what I learn, no hype.
Automating chaos only produces faster chaos: the five questions and the formula that decide what deserves automation and what does not.
Read article →
From Applications Engineer to presales that diagnoses: the unglamorous work behind a number a client made public.
Read article →
An API can return success without persisting the change. Twenty-three times the read-back caught it, and that is why there are nine gates and not one.
Read article →
Every refusal from the agent is an incident that never happened: the right question is not what your agent can do, but what it refuses to do.
Read article →
Frequency times minutes times people, divided by effort and risk. In a PDM with no bulk undo, risk is the term that decides.
Read article →
The name identifies, the metadata filters: the standard that survives is the one living in the CDE configuration, not in a PDF.
Read article →
I built an assistant that generates assemblies from a chat inside Inventor. This is what AI really does in CAD today, and what is still just a demo.
Read article →
Advertise, call, return and the gate. The protocol gives you three verbs; you write the fourth, and it is the one that decides whether the agent is safe.
Read article →
Thursday September 17, 4:30 PM, Las Vegas. A 60 minute Technical Deep Dive: what AU is, what I am bringing to the stage and what there will not be.
Read article →
Projects are not configured, they are born configured: folders, permissions, roles and ISO 19650 validated, from test hub to source of truth.
Read article →
Moving files is the easy part. States, properties, BOM and permissions need a bridge with dry-run, schema validation and audit.
Read article →
The agent has to go where people work, not the other way around: one logic and the same 9 safety layers in Vault Client, AutoCAD and Inventor.
Read article →
An LLM completes patterns: that is why it offers properties your Vault never had. The fix is not a better prompt, it is validating against the real schema.
Read article →
Years on the manufacturing side. Today I open the BIM front with no fake expertise: the same recipe of API, automation and limits, documented in public.
Read article →
The chat is just the front door. What defines an agent is its inventory of tools and the limits that govern every single one.
Read article →
This add-in does not contain a single prompt. And it is one of the projects that best explains how I decide when to use AI and when not to.
Read article →
When I say this in a meeting the same thing always happens: first silence, then someone starts making the list of tasks nobody wants to do.
Read article →
The manual audit of a 217-component assembly took about 20 minutes. Today it takes seconds. And speed is only half of the story.
Read article →
"AI for engineering" is the buzzword of the year. And a good part of what gets sold under that label does not survive 7 questions.
Read article →
Asking Inventor for an assembly in plain language is no longer science fiction. And the key piece has a name: MCP.
Read article →
The most common security mistake in an AI agent is not in the model. It is in the permissions it runs with.
Read article →
One of the projects I'm most proud of doesn't have a single line of AI in it. And that's exactly why it works.
Read article →
You can't just plug ChatGPT into Inventor and hope for the best. What actually works is a 3-layer architecture where the LLM never touches the CAD API.
Read article →
I built an AI agent with write access to Autodesk Vault in production. And the first thing I designed wasn't what it does. It was what it cannot do.
Read article →
"Can you build me an AI agent?" is the question I get the most. I almost always answer with another one: what problem are you actually trying to solve?
Read article →