The weekly strip · 102 issues · page 5 of 9
What shipped, and what to build with it
One post a week on the AI models and tools that matter to people shipping products: dates, prices, limits, and what I would actually use. Primary sources only.
- Issue 54
10 October 2026 · From the case files · 7 min read
Fraud signals in uploaded documents: what a model can and cannot see
A vision model reads a payslip well but never sees its PDF metadata. Which fraud checks belong in the model, which in code, with a Claude API example.
- fraud detection
- document processing
- Claude API
- structured output
- Issue 53
10 October 2026 · Building with AI · 7 min read
Cost per decision: a framework for pricing LLM classification at scale
A formula for classification cost: tokens, cache, batch, human review and error cost, filled with published Haiku 5.5, Sonnet 5.5 and Gemini prices, no invented accuracy.
- cost
- classification
- Haiku 5.5
- Gemini 3.8 Flash
- LLM pricing
- Issue 52
10 October 2026 · From the case files · 7 min read
OCR plus VLMs: document processing in 2024 and in 2026
The 2024 GPT-4 Vision plus OCR pipeline at 95% across 17+ document types, and what current Claude and Gemini document APIs change, with per-page cost maths.
- document processing
- OCR
- VLM
- Claude
- Gemini
- Issue 51
10 October 2026 · Building with AI · 6 min read
Multilingual AI features: serving 23 languages without 23 prompts
One prompt with a locale slot, language detection only where needed, per-language token costs and per-language evals. A design for 23 languages, not 23 prompt files.
- multilingual
- i18n
- LLM
- evaluation
- prompt caching
- Issue 50
10 October 2026 · Building with AI · 7 min read
PII and prompts: redaction patterns before the model sees the text
Redact in code before the model call: regex, NER and review, reversible placeholders, what GDPR and India's DPDP Rules imply, and how caching and retention interact.
- PII
- privacy
- redaction
- GDPR
- DPDP
- Issue 49
10 October 2026 · Building with AI · 7 min read
LLM guardrails: input checks, output checks and the ones that fire
Run cheap deterministic checks first, a small model screen second, and log how often each guard fires. A layered design with Haiku 5.5, structured outputs and Noul.
- guardrails
- safety
- validation
- structured outputs
- Claude Haiku 5.5
- Issue 48
10 October 2026 · Building with AI · 7 min read
Prompt versioning and rollback: prompts belong in git now
OpenAI is retiring prompt objects and Anthropic's Playground keeps no history. Version prompts like code, hash them into logs, and roll back without wrecking your cache.
- prompts
- versioning
- CI
- rollback
- prompt caching
- Issue 47
10 October 2026 · Building with AI · 7 min read
Cost observability for LLM apps: log usage per request from day one
Usage fields differ across Claude, OpenAI and Gemini. Normalise them, price each call with a dated rate table, and send cost and cache hits to Datadog per feature.
- observability
- cost
- Datadog
- LLM pricing
- Claude API
- Issue 46
10 October 2026 · Building with AI · 7 min read
Latency budgets for AI features: where the milliseconds go
Split an AI feature's latency into time to first token and decode time, then spend a budget on model tier, caching, output length and parallel calls, with a worksheet.
- latency
- performance
- LLM
- prompt caching
- TypeSafe Jev
- Issue 45
10 October 2026 · Building with AI · 7 min read
Streaming UX for AI answers: what to render before the first token
What users should see in the first second of an AI answer, how to render half-written Markdown without flicker, and how to cancel cleanly. Includes a React example.
- streaming
- UX
- React
- Markdown
- Claude API
- Issue 44
10 October 2026 · Building with AI · 7 min read
Function calling across Claude, GPT-6 and Gemini: one tool, three APIs
One lookup tool defined for Claude, OpenAI and Gemini: where the call comes back, how results go in, what parallel and strict mode look like, and a copyable adapter.
- function calling
- tool use
- multi-provider
- Claude Sonnet 5.5
- GPT-6
- Issue 43
10 October 2026 · Building with AI · 7 min read
Structured output without parsing: JSON schema modes vs System One models
Claude and Gemini schema modes guarantee the shape of a reply; TypeSafe's Jev returns typed answers with probabilities. What each documents, and when to use which.
- structured output
- JSON schema
- TypeSafe Jev
- Claude
- Gemini