The autonomous model factory.

The easiest way to own your AI models.

Facet builds, runs and improves a custom AI model for your product.

Open in

A custom model, from one message.

# eng-platform
M
Maya Okafor10:42 AM

@Facet every email our users get goes through a frontier model to decide if it needs a reply. It is most of our AI bill. Can you build a model for it?

{ }inbox-sample.jsonl12 MB JSONL
FacetAGENT10:43 AM

reply-check reads one email and returns whether it needs a reply, who it is from and what they ask.

  • Flags every email that needs a reply
  • Names the sender, never the forwarder
  • States the ask in one line
  • Never flags newsletters or receipts

Up to $9, about 20 min.

Open in Facet
Approved by Maya
FacetAGENT10:44 AM

Building reply-check v1. I'll post here when it's live.

reply-check v1 is live
Evals
4 of 4 passing
Charged
$7 of $9
Three weeks later
M
Maya Okafor9:12 AM

@Facet it missed a question buried in a forwarded thread. Those need a reply.

FacetAGENT9:13 AM

I added an eval: Catches asks inside forwarded threads. v1 gets 3 of 20 right on it.

The fix is 40 new examples and a retrain. Up to $3, about 6 min.

Open in Facet
Approved by Maya
FacetAGENT9:21 AM
reply-check v2 is live
New eval
20 of 20
Other evals
All held
Charged
$2 of $3
Acme/reply-checkPlannedRunningMergedPlannedRunningMerged

Train v1

Catch asks inside forwarded threads

v1v2
Quote up to $9, about 20 min4 of 4 evals passing, charged $7 of $9Quote up to $3, about 6 min20 of 20 on the new eval, charged $2 of $3
Evals
  • Flags every email that needs a reply24 of 25
  • Names the sender, never the forwarder25 of 25
  • States the ask in one line23 of 25
  • Never flags newsletters or receipts24 of 25
  • Catches asks inside forwarded threads new3 of 20
  • Catches asks inside forwarded threads20 of 20
Change 1
  • Build training data6m 40s
  • Train18m
  • Test on examples it never saw2m 10s
Change 2
  • Add 40 forwarded threads with a buried ask1m 30s
  • Retrain4m
  • Test on examples it never saw40s
inbox.py
client.chat.completions.create(
  model="reply-check",
  messages=[email],
)

Fully managed. Fully yours.

Built for one task.

A model that only knows your task can be small and fast.

You ask. Facet builds.

Approve a price and step away.

Evolves with your data.

Every correction makes the next version better.

Stop renting the intelligence your product runs on.

Frontier labs build one model for everyone. Facet builds yours.

Read the essay

Questions, answered.

What is Facet?

Facet, made by Halfspace, builds a custom AI model for a task your product repeats, such as reading every email or routing every support ticket. You describe the task in plain words and Facet does the rest. The model is small, so each call costs less than the same call on a large frontier model. You call it with the OpenAI SDK, and Facet keeps it right as you correct it.

What kinds of tasks is Facet for?

Facet is for narrow text tasks that run often, such as routing tickets, extracting fields, sorting text into categories, redacting personal data, rewriting in a house style, or grading text against your rules. Facet handles text today.

Do I need to write prompts or prepare examples for Facet?

No. A few real inputs help, but Facet writes the instructions and the examples itself. You describe the task and agree on what a good answer looks like, and you can open and edit anything Facet wrote.

Can my AI agent use Facet?

Yes. Facet has an MCP server at https://facet.halfspacelabs.com/mcp that works with Claude, Cursor, ChatGPT, Codex and VS Code, so your agent can build a model, ask for fixes and call it, while you approve anything that costs money.

How much does Facet cost?

Every account starts with $15 of free credit, once, with no expiry: it pays for drafting, training and calls. After that, calls cost from $0.04 per million input tokens, every build is quoted before it starts, and most first versions cost $1 to $3. Every rate is on https://halfspacelabs.com/pricing.

Can a Facet bill surprise me?

No. Every build shows its price before it starts, and that price is a cap it never goes past. A monthly spending limit, on by default, stops new work and calls when it's reached; Facet emails your workspace's Owners as spend nears it, and an Owner can raise it.

Can I take my model out of Facet?

Yes. Export it any time, free, with its model files and instructions in a standard open format, and run it on your own servers.

How do I get access to Facet?

Facet is open to invited teams first. Join the waitlist at https://halfspacelabs.com/waitlist with the task you have in mind, then sign in with the email we invite.

All questions