Try “KYC”, “Bedrock”, “Claude”, “insurance” or “ISO”.
Claude
Case study · Built with Claude
Skugal designed and engineered Pinpole, a product that turns a plain-English request into a cloud architecture, simulates its traffic and cost, and deploys it to the customer’s own cloud. Claude Code was our engineering environment, the Claude API powers Pinpole’s AI features, and Claude Desktop was our test client for Pinpole’s MCP server.
The challenge
Designing infrastructure usually means a diagramming tool, a spreadsheet cost model, hand-written Terraform and a load test that can only run after money has been spent. Mistakes in capacity or cost show up in production, or on the first bill.
What we built
Pinpole combines AI-assisted design, a visual canvas, a traffic and cost simulator, and keyless deployment in a single product.
Describe a workload in plain English and services, networking and connections stream onto the canvas. Users can also upload a whiteboard photo or diagram and have it converted into an editable architecture.
Drag-and-drop design across AWS, Azure, GCP and Cloudflare, with per-service configuration, live multiplayer editing, revision history and a connection rule engine that blocks invalid wiring.
A 100 ms tick engine with 30+ stateful service models (Lambda concurrency, SQS depth, DynamoDB capacity and more) forecasts latency, errors, bottlenecks and monthly cost before anything is provisioned.
A rule engine plus an AI pass suggests cost, latency and resilience fixes, and applies the chosen fix to the canvas in one click.
Pinpole acts as an OIDC issuer, so customers’ clouds trust it through federation and every deploy uses short-lived credentials. Generates Terraform, CloudFormation, CDK, SAM and ARM.
Everything in the UI is available through a REST API, TypeScript and Python SDKs, and an MCP server, so Claude and other agents can design, simulate and open canvases as tools.
Where Claude fits
We used Claude to build Pinpole, to power its AI features, and to prove that AI agents can use it as a tool.
Our primary engineering environment. Engineers used Claude Code across the codebase to implement the canvas, the per-provider service schemas, the simulation processors, the IaC generators and the cloud deployment integrations, and to write tests and refactor as the product grew. Every change went through human code review before merge.
Powers Pinpole’s AI features in production: generating architectures from natural-language prompts, extracting services and connections from uploaded diagrams, explaining designs in plain English, suggesting improvements after a simulation, and refining generated IaC. Model output is validated against Pinpole’s provider registry and rule engine before it reaches the canvas.
Our test client for the Pinpole MCP server (@pinpole/mcp). We used Claude Desktop to exercise the tools end to end, such as building an architecture, running a cost simulation and opening the canvas, from a natural-language conversation, and refined tool names, schemas and responses based on how Claude used them.
Architecture
Claude does the reasoning. Deterministic engines validate, simulate and deploy, and a person approves every deployment.
How we worked
The same delivery pattern we now use on client projects.
Product and architecture decisions, provider schemas and acceptance criteria written by our engineers.
Claude Code implements features against the spec inside the repository, with engineers steering and reviewing each change.
Automated tests, rule-engine checks on AI output and simulation runs against reference architectures before release.
Release, watch real usage of the AI Architect and MCP tools, and feed what we learn back into prompts, schemas and tests.
Responsible AI
Infrastructure changes cost money and can cause outages, so Pinpole never lets model output reach a cloud account unchecked.
What it shows
Pinpole was designed and engineered by Skugal. It is live at pinpole.cloud with a free tier, and it is the reference product for how we build with Claude.
Pinpole’s Model Hub lets teams choose the model behind AI workflows or bring their own keys. Claude is integrated through the Claude API, and the MCP server works with Claude Desktop, Claude Code and any other MCP-compatible client.
Yes. The same pattern applies to internal developer platforms, AI copilots over your own systems and MCP servers that expose your tools to Claude. We usually start with a short discovery and a fixed-scope pilot.
Next step
Tell us what you want to ship. We’ll come back with an architecture, a delivery plan and how Claude Code, the Claude API and MCP fit in.