# One layer to make inference smaller

Run large models on hardware you own, governed. This page shows how Tileward Models, Tileward Context, Tileward Documents and Tileward Governance work as one layer between your application and the model: it reduces what is sent to the model, controls the sources and tiles it can use, and lets you choose where every piece runs.

Page: https://tileward.com/platform/

## Two jobs

- **Reduce the cost (Models + Context):** compress model weights and stop resending the full conversation on every turn.
- **Control the work (Documents + Governance):** answer from approved files and refuse locked tiles before generation begins.

Use one piece or the full path.

## One question through the full path

A question leaves your agent carrying the whole thread. Governance classifies it and lets it through. Context collapses the thread to the turns that matter and pulls the relevant passage out of your documents. That small bundle reaches the model, the answer comes back with the passages it was given, and a line is written to the record. A question on a locked tile is refused at the gate and recorded, and nothing reaches the model.

- **In front of your agent:** one base URL, and every call your agents already make goes through it.
- **As a tool your agent calls:** connect the MCP endpoint, and the agent calls Tileward mid-conversation, keeping whichever model it already talks to.

28 clients have a recipe to paste in the [connect guide](https://tileward.com/docs/#connect), and anything else that speaks MCP or the OpenAI wire format works too.

## Request path: governed inside the operator's building

One request from a bank, hospital or ministry enters the operator's AI factory, in the operator's own country:

- First, a check: is the tile locked? If it is, the request is refused in about 14 tokens, before the model answers.
- If not, the model on the operator's GPU answers. It uses about a third of the memory, sees one page rather than the whole chat, and answers only from approved files.
- The answer goes back with its source passage, and every decision, allowed or refused, goes into an audit record kept for the customer's auditor.
- All of it runs on the operator's GPU, shared and billed per customer.

## Deployment boundary

The architecture stays the same; what changes is where it runs and whether Tileward ever receives your data.

- **Hosted API:** fastest start. Tileward operates the service.
- **Your VPC:** the data path runs inside your cloud account.
- **Air-gapped:** the full stack runs on your hardware with no Tileward data path.

## Bring one real workload

In thirty minutes an engineer maps the path, runs the relevant eval on it, and shows where the savings and the controls come from. [Book 30 minutes](https://calendar.app.google/hg2DzFTxKJgwFWm19), [start free](https://app.tileward.com/?join=explore), or email hello@tileward.com.

## Related

- [Tileward Models](https://tileward.com/models/): run larger models on existing GPUs
- [Tileward Context](https://tileward.com/context/): send only the conversation that matters
- [Tileward Documents](https://tileward.com/documents/): answer from approved files, with citations
- [Tileward Governance](https://tileward.com/governance/): refuse locked tiles before inference
- [Security](https://tileward.com/security/): compare the security boundaries
- [Evals](https://tileward.com/evals/): what we measured