How we use AI

AI proposes. Something checkable decides.

Claude is the backbone of how we build. These are the models and checks behind the six products.

The stack

Claude (Opus 5.5, Sonnet 5.5, Haiku 4.5)

Engineering partner through Claude Code, plus research synthesis, review and structured extraction. The core of how we build.

GPT-6 “Astra” on AWS Bedrock

Bulk authoring and independent review, behind a budget runner with a hard cap, a ledger and a kill switch.

Open models on Bedrock (Qwen3, Qwen3-VL, Gemma, GLM-5)

Reasoning, image-to-ingredient detection and translation, chosen for cost and speed.

Image models (Stable Image on Bedrock)

Card and deck art for Astral Nexus, with no AI pictures of deities.

Sarvam and DeepL

Translation for Indian and European languages in ConsumerSuit.

Deterministic checks

ngspice, legislation.gov.uk, an open-source ephemeris, and our own gates. The AI proposes; something that can be checked decides.

Working agreements with the machine

Claude Code is the engineering partner. The founder sets what is built and why; Claude implements, tests and reviews. The journal attributes ideas and engineering separately.
A simulated board challenges every decision. Nine named department heads (engineering, AI, security, QA, design, growth, finance, legal, data) each report from the real codebase, and worker agents must have their claims verified against it.
Gates run after the action, not before. A rule that needs a correct choice in the moment gets broken. So we write the script: verify before saying done, rights gate before any story ships, budget runner before any paid call.
Spend is capped and logged. Paid models run through a runner with a hard ceiling, a ledger and a kill file. Cost per analysis in ConsumerSuit is about one and a half US cents.
Failures are recorded. A mistakes archive, dead-ends file and pre-registered experiments keep us from repeating or hiding them.

What this has produced

Speed with receipts

Six products and 1,700+ commits in about four months by one founder, each with tests and a journal.

Honest measurement

Determinism receipts, scorecards with their own answer-key bugs found, and a failed pre-registered test acted on.

Grounded outputs

Citations from statute text, numbers from a solver, astronomy from a library. The model never invents the number.

Why this matters for Claude: we are a UK startup that depends on Claude for the core of how we engineer, research and review, and that has built discipline around using it responsibly.