service · AI & MLClaude-native deliverybuild + operate · on-call

Claude Code consulting from a studio that operates what it ships.

Claude Code consulting, principal-only. Claude Code is our engineering surface — we build on Claude by default, write the evals on day one, ship in two-week cycles, and stay on the pager when it is live. No juniors on the keyboard. No demo theatre.

claude code consultingclaude-nativeprincipal-onlybuild + operateevals day one
01The shift · what Claude Code changes

It moves the hard conversation to design time.

Claude Code is not autocomplete. It reads the repo, drafts the plan, and acts on the real system through MCP — so review happens before merge, not after the incident.

The codebase is the prompt

Claude Code reads your repo, not a snippet. With a 1M-token context window it holds an entire service in view — call sites, types, tests, and conventions — so changes land in your idioms instead of a generic stub.

Review moves earlier

Architecture memos, diffs, and migration plans are produced and read before a line is merged. The expensive conversation happens at design time, where it is cheap to change your mind.

Tooling is first-class

Through the Model Context Protocol and the Claude Agent SDK, Claude Code talks to your build, your test runner, and your internal services directly — so it acts on the real system, not a description of it.

Velocity without ownership drift

Output multiplies, but the senior engineer still signs the PR. We treat Claude Code as leverage on judgment, never a replacement for it — the human who can defend the change is the one who ships it.

02How we work · build + operate

We don't hand you a Claude demo. We operate it.

A principal in every session, graders before features, and an SLA-backed pager after launch. The same people who build it run it.

build

Architect with Claude, ship with humans

A principal drives the session — scoping, reading diffs, rejecting the lazy answer. Claude Code drafts; the engineer who owns production approves. Every PR is defensible on a call.

evals

Graders before features

We write a golden set and LLM-as-judge graders on day one. No Claude-touching path ships without a pass/fail number, and the suite runs on every PR and nightly so a model update can never silently regress you.

operate

We stay on the pager

When the system is live, we keep the runbook current, tune prompts and graders as the world changes, and answer the page under an SLA. You inherit a system you can run — and a team that already runs it.

03Deliverables · what shows up in your repo

Four things, all yours, all versioned.

01

Claude integration

Typed calls to the Messages API or Bedrock/Vertex, prompt modules, fallback policy. Versioned like code.

src/claude/*.ts
02

MCP + tools

Model Context Protocol servers wiring Claude to your build, data, and internal services with typed schemas.

mcp/*.ts · tools.ts
03

Eval harness

Golden sets, LLM-as-judge graders, PII and jailbreak guards, regression runs on every PR.

evals/*.jsonl · ci.yml
04

Runbook + on-call

What to do when latency spikes, a grader red-lines, or a model update lands. Plus the pager rotation.

docs/runbook.md
04In the chair · a real Claude Code session

A change with a grader attached.

Plan, edit, eval, commit — with a hard stop for human review before anything merges. This is the loop on every engagement.

bashclaude code · refunds · idempotency
$claude "add idempotency keys to the refund endpoint, with a regression eval"reading repo · 41 files in context (api/, db/, evals/) ...plan: dedupe on Idempotency-Key header → ledger lookup before charge reversaledit  api/refunds.ts          +38  −6edit  db/migrations/0042.sql  +12  −0edit  evals/refunds.jsonl     +9   −0   (replay + double-submit cases)$npm run eval -- refundsgraders: 9/9 pass · double-submit blocked · no double refundnote: principal review required before merge — diff attached to PR$git commit -m "refunds: idempotency keys + eval"committed 8f3a1c2 · CI green · awaiting human approval
05Proof · eval discipline

If we can't grade it, we don't ship it.

Claude is non-deterministic; our delivery is not. Every Claude-touching path lands with a measurable bar that runs in CI.

golden sets

Real inputs from your domain, labelled with the answer you would defend in a post-mortem.

llm-as-judge

Claude grades Claude against a rubric you sign off on — so quality is a number, not a vibe.

guards

PII, jailbreak, and policy checks run inline. A failed guard fails the build, not production.

regression

The full suite runs on every PR and nightly. A model swap is a code change with a test, not a leap of faith.

The same discipline runs deeper on our AI workflow automation work, and we apply it inside compliance boundaries on sovereign AI for the GCC.

06More · Claude-native work

Where to go next.

07Questions · Claude, specifically

What engineering leaders ask before they pick a Claude studio.

Which Claude models do you build on?

The current Claude family — Opus 4.8, Sonnet 4.6, Haiku 4.5, and Fable 5 — chosen per task: a capable model for hard reasoning, a fast one for high-volume paths. All ship with a 1M-token context window, and we write the integration so the model is a swappable dependency, not a hardcode.

Is Logic Layer an Anthropic partner?

No. We are a Claude-native studio: we build on Claude and Claude Code by default and specialise in them, but we are not a member of the Claude Partner Network and make no claim to an official Anthropic relationship. If you need the Partner Network specifically, that program is run by Anthropic directly.

Can you deploy on AWS Bedrock or Google Vertex AI?

Yes. Claude is available on AWS Bedrock and Google Vertex AI as well as the Anthropic Messages API. We deploy into whichever your cloud, compliance, and data-residency posture requires, and keep the integration portable across them.

How is our data handled?

Under your DPA and your key management, in your cloud account unless you explicitly ask us to host. Prompts and outputs stay inside your boundary; we wire PII guards into the eval harness so sensitive data is caught before it reaches a model.

How do you govern Claude in production?

Every Claude-touching path has graders, guards, traces, and cost and drift alarms wired to your observability stack. Prompts and tool schemas are versioned in your repo and reviewed like code, so there is an audit trail for every change a model can make.

Who owns the code?

You do, from day one. The repo, the prompts, the MCP servers, and the eval suite are yours. We leave you a system your own engineers can read, run, and extend without us.

What does "operate" actually cost?

A monthly retainer tied to eval runs and on-call, reviewed quarterly — qualitatively a fraction of the build, scaled to how much you lean on the system. No surprise hours, and you can take operations in-house whenever your team is ready.

end · next step

Send us one repo.We'll send back a plan in 48 hours.