Guides
Written to be
even if you never hire anyone.
Two kinds. Buying guides for when there is a five-figure decision open and nobody will give you a straight number. Diagnostics for when something is already broken and you need to know which layer it is in.
Each one is written so you could act on it without us — including the parts where the honest answer is not to build the thing, or not to hire anyone at all.
Buying guides
For a decision that is open right now.
What an AI feature actually costs to build
Real engineering hours, real inference costs at real volumes, and the line items that surprise people. Including when the honest answer is not to build it.
“I have to put a number in front of my board and I don't know if it's $10k or $200k.”
OCR or an LLM for invoice extraction?
Traditional OCR is brittle on layout. LLMs invent numbers. They fail differently, and which failure you prefer decides the architecture.
“Our OCR breaks on new layouts and the LLM makes things up. Which problem do I want?”
Build or buy document extraction
When a vendor API is obviously right, when it obviously isn't, and the three questions that decide it. Most teams get this backwards in both directions.
“Do we use a vendor or build our own? Getting this wrong costs a quarter.”
RAG or fine-tuning?
They solve different problems and are constantly compared as if they were alternatives. A decision procedure, and the case for doing neither.
“Everyone gives me a different answer and both are expensive to get wrong.”
Hiring an AI engineer vs bringing in a firm
A loaded salary, four months of hiring, and a market where you cannot evaluate the candidates — against a contract. An honest comparison from someone who sells one side.
“$180k a year and four months of searching, or a contract. Which?”
Diagnostics
For a system that is already in front of customers and misbehaving.
Why your RAG system returns wrong answers
Six causes, the symptom each produces, and how to tell them apart in an afternoon. It is almost never the model.
“The chatbot is confidently wrong in front of customers and I need it fixed.”
How to evaluate an LLM pipeline
Building a golden set, choosing metrics that mean something, and gating deploys on them. The smallest version that works, buildable in a week.
“I know we should have evals. I have never built one and don't know where to start.”
When extraction accuracy collapses in production
It worked on your test set and fails on real documents. The distribution gap, how to measure it, and what to do that isn't 'try a bigger model'.
“It works in testing and falls apart on real customer documents.”
When the model bill triples and nobody knows why
How to attribute spend to a feature, the four things that usually dominate, and the changes that cut cost without cutting quality.
“The bill tripled and nobody can explain which feature did it.”
Adding provenance to AI outputs
Compliance asked why the system said that, and you cannot answer. What to record, when to record it, and what it costs to retrofit.
“Legal or an enterprise customer just asked 'why did it say that?' and we can't answer.”