Field notes
What we think.
Notes on enterprise AI from the delivery side.
Governed RAG on a medallion architecture
A retrieval system inherits the access rules of the table it reads. This pattern keeps those rules attached to the data through every layer, from ingest to the vector index.
ReadWhy AI programs stall after the pilot
The prototype is rarely the constraint. Programs stall on identity, data access, and the absence of a team that owns the system in production.
ReadDon't automate the workflow you have
The process you are about to accelerate was shaped by the limits of the people who ran it. Those limits have changed.
ReadChoosing a model is the last decision, not the first
Most of the quality in a working system comes from the components around the model. That is also the case for staying vendor neutral.
ReadBuilding multi-agent systems with the OpenAI Agents SDK
Handoffs, guardrails, and tracing are the three primitives that decide whether a multi-agent system can be operated. Here is how we wire them, and what we have learned about where each one belongs.
ReadWhat a production AI system actually costs
Token pricing is the smallest line in the bill and the only one most estimates contain. Here is the unit that matters, and a worked example of what it exposes.
ReadAIoT: inference where the network is not a given
A plant does not stop because a link to a region is down. That single constraint decides most of the architecture for industrial AI.
ReadEvaluations that gate a deploy
An evaluation suite that nobody can fail is a dashboard. This is how we build one that blocks a release, and what it costs to keep it honest.
ReadPrompt injection is an authorization problem
You cannot prompt your way out of prompt injection. The fix is to stop giving the model authority it can be talked into using.
ReadThe control plane an agent estate needs
One agent is a project. Twelve agents is an estate, and an estate needs identity, policy, and observability that do not live inside any of them.
ReadChunking is a retrieval decision, not a preprocessing step
Most retrieval quality problems we are called in for are chunking problems wearing a model costume. The fix is upstream of anything you can prompt.
ReadHuman in the loop is an architecture decision
Review everything and you have built a slower version of the old process. Review nothing and you have built a liability. The gate belongs on the effect.
ReadGetting a pilot through change control
The gap between a working prototype and a deployed system is mostly paperwork you can prepare for. Most teams meet it for the first time at the end.
ReadThe metrics a steering committee will accept
Model accuracy does not survive contact with a business review. These are the measures that do, and the baseline problem that undermines most of them.
ReadBring us the problem.
Tell us the outcome you are trying to create, what you have already attempted, and where the constraints are.
Contact nuperX