Your documents · your model · your weights

Your documents,
trained into your own model.

Scalty reads what your team already wrote, writes training examples from it, fine-tunes an open-weight model and measures it against the base — then serves it behind an endpoint you own.

No card required · export your weights any time

studio.scalty.com / customer-support / overviewTODO

Customer Supportiteration 7 · stage 3 of 43 running

01 DATATickets 2024952 documents

02 DATASETSupport pairs #61 240 pairs

03 TRAINsupport v762%

04 DEPLOYapi-prodserves v6

Active work3 jobs

Training · support v7Training62%14:32

Reading · upload 7c21Reading40%03:05

Deploy · api-prodWarming—01:12

POST/api/v1/chat/completions200312 req/minp95 840ms

The loop

build on the left · serve on the right · repeat

Documentsmarkdown · plain text

Datasetquestions and answers written from them

TrainLoRA on open weights

Compareheld-out, base vs yours

ServeOpenAI-compatible endpoint

Meterevery minute on the ledger

What it does

From a folder of documents to a model that answers like your team.

Documents become training examples

Upload Markdown and plain-text files. Training reads them and writes the question–answer pairs it learns from; you start one run, not a preparation step.

Trained, not searched

LoRA on a license-clean open-weight base. Settings are derived from your data or set by hand; time and price are estimated before the run starts.

See what it learned

Held-out questions from your own documents, answered by the base and by your model, side by side. The model card reports what was measured.

Memory for facts that change

Training carries voice and terminology. For prices, rules and stock, attach a memory collection — answers come back with a citation into the source.

Three ways to use it

Answer in the playground, call an OpenAI-compatible endpoint with a key, or download the weights and run them on your own machine.

Metered, on a ledger

Every figure on the billing page comes from the usage ledger. $5 of credit every month; no plans, no seats.

API-first

Point the OpenAI SDK at your endpoint.

Streaming, usage and citations come back in the shapes you already handle. Keys are scoped to the organisation; every deployment has revisions you can roll back.

chat/completions · predictions
API keys per organisation, metered requests
export-manifest: take the weights with you

studio.scalty.com/api/v1TODO

from openai import OpenAI

client = OpenAI(
  base_url="https://studio.scalty.com/api/v1",
  api_key="sk_live_…",
)

r = client.chat.completions.create(
  model="support-assistant-prod",
  messages=[{"role": "user", "content": "Hi"}],
)

200stream · 41 tokensfirst token 180mssupport-assistant-prod · v6

PricingTODO

Metered. Read straight off the ledger.

No plans, no seats. A run shows its cost while it works and its billed minutes when it ends; the balance page says exactly what was spent. The rates shown are today's rate book.

$5usage credit every monthno card to start · set your own overage limit

TrainingLoRA fine-tuning, per accelerator minute$0.05/ min+ $0.25 per launch

Tokenstext inference$0.20/ 1M in$0.90 / 1M out

Teacher worktraining data preparation$1.00/ 1M tokens

Embeddingsearch preparation$0.02/ 1M tokens

Predictionstabular classification and regression$1.00/ 1 000 rows

Storagedocuments, datasets, artifacts$0.10/ GiB · month

Bring one folder of documents. Leave with a model that is yours.