Blog

Working detail from systems that actually shipped.

Notes from the build rather than commentary on the industry. Each post carries something specific — a real decision, a real failure, a real number — because a post without one of those is just restating what a model already knows.

90 posts, grouped below by what they are about.

01

Latest

Agents

Freeze the evidence, or you are not measuring the agent

Cloudflare's new security harness runs no AI in its front half, and the reason it gives is an eval reason: a versioned evidence snapshot is what separates retrieval drift from reasoning drift. It is also what makes the parallel fan-out cacheable.

Retrieval & eval

Embed the page, not the OCR text

EmbeddingGemma 2 embeds a page image in 280 tokens and matches it against text in one space. Indexing a million scanned pages is $3.36 — and the 6× storage cut costs 18.83% of visual retrieval.

Agents

A monitor is a rate, not a gate

A paper measures how often nine frontier models slip a protected value past a monitor while genuinely trying to be helpful. The rate is 0.9%, and 0.9% is a certainty at a thousand episodes.

Cost & ops

The search line is 3% of the answer line

AI Search is GA with published rates and billing from 1 November 2026. Indexing a document is $0.00146 and a semantic query $0.00075 — the line that changes a decision is the one nobody is looking at.

Retrieval & eval

Retrieval without a k is a bill you cannot write down

PageIndex indexes a document for $0.001 a page and lets an agent reason down a tree instead of measuring cosine distance. The index is the small half. Nothing caps the number of steps, and that is the bill.

Extraction

The task mix decides the review bill

A DOCX editing benchmark publishes eight per-task-group pass rates instead of one score. Read as a workload rather than a score, the same product costs $0.155 or $0.211 a document.

02

Extraction

The task mix decides the review bill10 min Masking columns is not masking data9 min Document-level confidence is nearly useless7 min 03/04/2026 is two different dates7 min Teach the extractor to refuse6 min 1.234,56 is not 1.2347 min Page two assumes you read page one8 min The review queue is the product8 min Your golden set is your happy path8 min Stop parsing JSON out of prose7 min What a scanned page actually costs8 min The PDF says it has text. The text is wrong.8 min The same invoice, uploaded three times7 min 97% accurate is not a number8 min You changed the prompt. Now what about the last 400,000 documents?8 min The row that starts on page two7 min Where traditional OCR still wins8 min Fail open or fail closed — pick before you ship7 min The field you assumed was printed7 min Straighten the page before the model sees it7 min Mapping columns you have never seen before8 min A schema that fails loudly beats a blob that fails silently7 min

Feeds

AI project rescue →

22 posts

03

Retrieval & eval

Embed the page, not the OCR text11 min Retrieval without a k is a bill you cannot write down10 min Reranking is 86% of the latency and 15% of the gain10 min Why your vector search returns irrelevant results8 min Hybrid search beats a bigger embedding model8 min A retrieval eval you can build in a day9 min When changing RAG chunk size will not improve accuracy8 min When a reranker is worth the latency7 min RAG citations: attributing an answer to a passage, not a document8 min The RAG bug that shows tenant A tenant B's data8 min Filter first, then search7 min Do I need RAG? Usually not yet7 min How to update embeddings when documents change7 min Perfect retrieval, wrong answer8 min Tables are invisible to semantic search7 min Where the milliseconds actually go7 min Rewriting the query, without losing the question7 min How to change an embedding model without downtime8 min

Feeds

RAG and retrieval →

18 posts

04

Agents

Freeze the evidence, or you are not measuring the agent11 min A monitor is a rate, not a gate11 min Preview isolation stops at the binding10 min Running out of disk in the middle of a batch9 min How many simulated conversations before the number means anything10 min An unattended agent run has no ceiling unless you build one9 min How to stop an AI agent doing the same action twice7 min What to do when an AI agent's approval never comes7 min How to restrict which tools an AI agent can use7 min Retry is not resume7 min How to set a spending limit on an AI agent run7 min What to record in an AI agent audit log7 min Read the retry-after header6 min Why your AI agent gets stuck in an infinite loop7 min Why every AI agent needs a dry run mode7 min Running an MCP server in production8 min What the approver needs on screen7 min How to add human approval to an AI agent7 min

Feeds

AI integration →

18 posts

05

Cost & ops

The search line is 3% of the answer line11 min The 58.3% is the median. The bill is the mean.10 min A cache break does not raise an error. It raises the bill.10 min The same 2,500 calls, two loop orders, half the bill10 min Failover is sticky. The cache is not.9 min Which feature tripled the bill8 min Prompt caching, with the arithmetic7 min The price rise already on the calendar6 min Output tokens are five times the price6 min Half price, if you can wait7 min You are paying for every failed call6 min Try the cheap model first8 min Paying to re-read the same tokens7 min Embedding a million documents7 min Quoting a per-document price you can defend8 min Streaming costs the same and feels twice as fast6 min

Feeds

What an AI feature costs →

16 posts

06

Agency & white-label

Questions to ask before hiring an AI developer8 min White-label AI work, and how it actually goes8 min How to break an AI demo in ten minutes7 min Scope it as a measurement, not a feature8 min The eval set is the deliverable7 min Why your AI demo works and production fails7 min Fixed-scope audit or retainer: which to buy first7 min An AI project handover checklist8 min

Feeds

White-label AI development →

8 posts

07

Build notes

Two clocks6 min A rule pack, not a model7 min Detecting a contact form that goes nowhere7 min Plain text, one at a time, from your own Gmail7 min Agent state as plain files7 min The prospect list never leaves the machine7 min I committed a .env to a public repo and did not notice for seven months8 min Why my crawler is deliberately slow6 min

Feeds

Case studies →

8 posts

Something here

the audit is the cheapest way to find out for certain.

Book a 30-minute call →Start with the $2,500 audit