Claude reasons. NVIDIA runs it. We ship the product.
Three products are launching on that stack. ACCA is the proof — Kontai reads the document, Claude decides the booking, and NVIDIA hardware serves the model. SparkVibe and Artizai ship the same way.
CODES AI is joining both startup programmes. The products are already built for the stack those programmes exist to support: Claude for the judgement, NVIDIA hardware for the compute. ACCA is where that is easiest to see.
Claude
Anthropic
Claude Startups
The judgement layer. Claude is the model that reads a document, a codebase or a contract and says what it decided.
01
Opus, Sonnet, Haiku
Opus 5.5 for the hard reasoning. Sonnet and Haiku when the same work should be fast and cheaper. One family, picked per job.
02
A trail, not a guess
Long context holds a ledger, a repository or a pile of invoices. Tool use lets the model act. If it is not sure, it stops.
03
Claude Startups
Anthropic’s programme for companies building on Claude: API credits, Claude Team, office hours with Applied AI, and the Startup Stack around the product.
The hardware layer. The models only stay fast if the GPUs, the CUDA stack and the inference path are real.
01
GPUs and CUDA
Training and inference on NVIDIA accelerated computing — the hardware ACCA’s document and booking models actually run on.
02
NVIDIA NIM
Inference microservices for serving foundation models without standing up a research cluster first. Secure, repeatable, production-shaped.
03
NVIDIA Inception
The startup programme: SDKs, Deep Learning Institute training, preferred hardware and software pricing, and a path into the NVIDIA ecosystem. No equity taken.
This is the path inside Kontai on acca.codes-ai.uk — Claude decides, NVIDIA serves, the ledger keeps the reason.
01
Document in
A PDF, a photo or an e-invoice lands in ACCA.
02
Claude reads
Amount, date, tax and counterparty — each value stays attached to the document.
03
Claude decides
The tax case and the account, with the reason written down. Unclear means flagged.
04
NVIDIA serves
That inference runs on NVIDIA hardware, not a shared guess in the browser.
05
Books close
A balanced entry. The trail is still there a year later.
Launching
Three products. One stack.
ACCA is the full demonstration. Claude does the judgement. NVIDIA hardware does the compute. The booking keeps the reason.
01 / 03 · acca.codes-ai.uk
ACCA
Kontai. The books finish themselves.
ACCA is Kontai — the live platform on our subdomain. A document comes in. Claude reads the amount, the tax case and the counterparty, then proposes a balanced booking with the reason attached. What is not clear is flagged, not guessed. NVIDIA hardware serves that inference.
Reads the document, chooses the tax case, and writes the booking a reviewer can follow.
NVIDIA
Document and booking models run on NVIDIA accelerated hardware.
Claudetax case reasoned
NVIDIAinference live
Ledgerdebit = credit
02 / 03 · vibe.codes-ai.uk
SparkVibe
The studio that ships with Claude.
SparkVibe turns repositories and databases into plain-English docs, knowledge packs and SDLC-ready context for Claude and Cursor — a Context Vault, a database explorer, and a prompt lab, so the model starts from the product instead of a blank chat.
Holds the codebase, the schema and the prompt history Claude and Cursor actually use.
NVIDIA
Evaluation and heavier model runs sit on NVIDIA compute, not a laptop fan.
Context Vaultrepo in plain English
ClaudeSDLC context ready
NVIDIAevals on GPU
03 / 03 · artizai.uk
Artizai
An AI platform, served on NVIDIA.
Artizai is the CODES AI platform for product intelligence: Claude for the reasoning a user actually sees, NVIDIA for the serving layer underneath, and the same shipping discipline we use on ACCA and SparkVibe.