# Specialist Factory

Connect an agent to train **task models** — small open-weight models on proprietary outcomes for one scoped job — then serve them under an existing harness.

## Positioning (do not confuse)

- **Brand:** Specialist Factory
- **Product:** task model (weights under the agent)
- **Not:** a “specialist agent.” Agent/harness calls the task model.

Hero: Your LLM is your agent’s biggest weakness. Put a task model under it. For basically free.

## Base URL

Control-plane origin that served this file (e.g. `http://127.0.0.1:3000`).

## Product surface

| Route | Purpose |
|-------|---------|
| `/` | Beachhead landing |
| `/pricing` | Plans + Stripe Checkout |
| `/trust` | Trust ladder |
| `/how-it-works` | Stack + steps |
| `/studio` | Studio overview |
| `/studio/new` | Create task model |
| `/studio/projects` | Projects |
| `/studio/billing` | Subscription |

## Trust tiers

| Tier | ID | When |
|------|-----|------|
| Shared cloud | `tier0_saas` | Pilots |
| Your cloud | `tier1_byoc` | Default for proprietary data |
| Confidential GPU | `tier2_confidential_gpu` | Attested TEE |

Tier 3 on-prem is out of scope — do not offer it.

## Flow

1. `GET /api/recipes`
2. Confirm dataset: `features`, `label`, optional `amount` / `reward`
3. Choose trust tier (default `tier1_byoc` for proprietary data; `tier0_saas` for pilots)
4. `POST /api/projects` `{ name, trustTier }`
5. `POST /api/datasets` — URI only for BYOC/CC; never paste plaintext rows into chat
6. `POST /api/jobs/eval` — baseline vs frontier
7. `POST /api/jobs/estimate`
8. `POST /api/jobs/train`
9. Poll `GET /api/jobs/:id`
10. If confidential GPU: `GET /api/attestation/jobs/:id` — stop if not verified
11. `POST /api/serve` — point the user’s agent at the endpoint

## Billing

`POST /api/billing/checkout` `{ planId: "pilot"|"build"|"scale"|"enterprise" }`

## Example

`/fine-tune-and-serve I have a review agent on a frontier LLM for payments risk. Help me train a task model on our labels, eval vs frontier, and deploy under the agent. Prefer BYOC.`