API World 2026
Designing an API Layer for AI Providers You Can’t Trust
September 1 to 3, 2026 · Santa Clara Convention Center, Santa Clara, CA
Every model provider will eventually rate-limit you, deprecate the model under you, change its refusal behaviour overnight, or return something that parses but is wrong. This is the layer I put between my products and all of them, and what it cost me to learn each rule.
Full talk · 26 min
What the talk covers.
- Why provider failure is a normal operating condition rather than an incident, and what that changes about your error handling
- The difference between a response that failed and a response that succeeded while being useless, and why only one of them shows up in your metrics
- Retry, fallback and degradation policies that do not quietly spend your budget
- Where to put the seam so swapping providers is a config change rather than a rewrite
If you were in the room.
The quickest next step is a short call. Tell me what you are building and who it talks to, and I will tell you straight whether I am useful to you. If you would rather start with something concrete, the Voice & Trust Audit is $4,500 and takes 2 weeks.
Short clips.
Behavioural drift: changing the model changes the product
Newer is sometimes better. It should never be the default assumption. 3 min
Judge every AI request on risk before a model is called
Each path needs an approved fallback, decided before anything goes wrong. 2 min
Keep up
Follow along on LinkedIn.
New talks, what I am building, and what I learn running AI in a real business. It is the best place to keep up, and to say hello.