Question 1

What does CloudCrane actually build?

Accepted Answer

Data pipelines that turn thin, messy records into structured data your AI agent can call as a tool, and the products that put it to work: semantic search, agents and recommendations. Menus, product catalogs, exercise libraries, labels — anywhere a field has to be right. Shipped examples: 1,324 exercises enriched for an injury-aware coach (Zabber), every dish tagged against 14 allergen classes (Gobbles).

Question 2

Is CloudCrane a product or a services company?

Accepted Answer

Both. We design and build custom pipelines, and the core guardrails now run as a platform in early access: sign in with Google, import a file, curate it, publish it, and connect the result to your agent as a tool. The Product page shows exactly what's available today and what's still planned.

Question 3

What does early access include?

Accepted Answer

Sign in with Google and you get a workspace straight away: import a table (CSV, TSV, JSON, Excel or pasted), define the values each field may take, preview a run, review what needs a person, correct what's wrong, publish a release, and connect it to your agent over MCP or REST with a key you can limit or revoke. During early access each workspace can hold 1,000 records and make 500 AI calls a month. Values settled by rules, records that haven't changed, and text the pipeline has already seen don't use AI calls. If you need more, get in touch.

Question 4

How does my AI agent use the data?

Accepted Answer

You publish a release and deploy it as a tool. Agent frameworks connect over MCP; anything else calls plain REST. The tool's schema is built from your contract, so the filters and allowed values are the ones you defined, and anything else is refused before a query runs. Exclusions (no dairy, nothing that loads an injured shoulder) are applied in the database query, and a record whose value isn't known yet is left out rather than assumed safe. Each row comes back with a receipt, so the agent can say where an answer came from. Each key can be given its own rate, monthly cap, expiry and permanent exclusions, and revoked on its own.

Question 5

Is my data private?

Accepted Answer

Each workspace is isolated in the database itself: every table that holds your records is protected by row-level security, so one workspace can't read another's, even if a query has a bug. The AI step only receives the columns you choose for a field, and you can export your data as CSV at any time.

Question 6

How long does a typical pipeline project take?

Accepted Answer

It depends on scope. Small automations typically take 2-4 weeks, medium-sized pipelines take 1-3 months, and large-scale enterprise data infrastructure runs 3-6+ months. We'll give you a realistic estimate after understanding your data.

Question 7

What kind of AI enrichment can you add?

Accepted Answer

Classification into fixed vocabularies (difficulty, allergens, cooking method), extraction from menu photos and free text, safety tagging like injury risks, short summaries, and embeddings for semantic search. Every output is validated against the schema, and anything outside the allowed values is dropped.

Question 8

How do you keep AI output trustworthy?

Accepted Answer

Facts that don't need a model are computed in code. AI output is checked against the contract you wrote. Safety rules are enforced by the system rather than trusted to the model: allergens can only ever be added, never removed, and when a tool is asked to exclude something the filter runs in the database query, so a record that hasn't been tagged yet is left out rather than assumed safe. Agents are tested with scripted scenarios, backed by 2,000+ automated tests across the two products we shipped.

Question 9

What's your tech stack?

Accepted Answer

Mostly TypeScript and Node on the backend, Postgres with vector search, React on the web and Flutter on mobile, with CI on every change. We deploy wherever your data already lives and pick tools per project rather than forcing one stack.

Question 10

How do we get started?

Accepted Answer

Two ways. To try the platform, sign in with Google and import a file. For a custom build, use the contact form below: we'll ask about your data sources and what a wrong value would cost, then follow up on scope, timeline and next steps.

Data you canstake safety on.

Any catalog.Tagged against your rules.

Where CloudCranesits.

Then your agent calls it

Demos are easy.These rules hold up.

Runs that survive crashes

Never enrich the same dish twice

Allergens can only be added

Safety isn't left to the model

AI features tested like software

Embeddings that mean something

Questionswe get asked.

Can't afford a wrong tag?Make every field provable.