Quick Start — Live in 15 Minutes

This guide takes you from zero to a working, monitored model endpoint you own, in about 15 minutes. It assumes you have a k3ld account and the web console open at model.k3ld.com.

What you'll end up with

• A connection to your data (or a bundled sample dataset)
• A fine-tuned model in your brand/domain (we'll use a chat assistant as the example)
• A live endpoint you own — hosted by k3ld (default) or in your own AWS account (BYO)
• A drift monitor watching it

Step 0 — Sign in (1 min)

1. Open https://model.k3ld.com.
2. Sign in with your org credentials.
3. You land on the Build stage.

Step 1 — Connect a data source (2 min)

1. In Build, open the data sources view and click add Add connection.
2. Choose your source type: S3, GCS, or a warehouse connection.
3. Provide the bucket/connection details. The platform reads your source directly — no export or upload needed.
4. Click save. The connection appears with its table count.

No data yet? Skip this and use the bundled example dataset in Step 3 — fine-tuning works the same way.

Step 2 — Explore the catalog (2 min)

1. In Build, open the catalog (Explore) view.
2. Your connected tables appear in a collapsible catalog tree.
3. Click any table to see its lineage — where it came from and what consumes it.

This is your unified catalog. Nothing was moved or copied.

Step 3 — Build: fine-tune a model (5 min)

1. Click Build → extension Fine-tune.
2. Pick a use-case preset. For this quick start choose chat Chat assistant (messages) — it pre-fills a base model, a bundled example dataset, and sensible hyperparameters.
3. Data: keep the bundled chat-assistant.jsonl (OpenAI-style {"messages":[...]} rows) or upload your own.
4. Read the cost estimate shown for the run (a fixed infrastructure cost, not per-token).
5. Optional but recommended: tick "Run GPU smoke before launch" — the platform does a short GPU test of your data + config before spending on the full run.
6. Click Launch.

The run appears in Build. Completed runs show a confidence indicator (based on how similar past runs performed), so you know at a glance how likely this exact config is to succeed.

⏱️ A small fine-tune on a g5 instance typically completes in 10–40 minutes. You can close the tab; it finishes in the background and notifies you.

Step 4 — Ship: deploy your model (3 min)

1. Click Ship → rocket_launch Deploy model.
2. In the Model dropdown, pick your completed fine-tune run. The Model S3 URL auto-fills.
3. Choose an instance type (the estimate shows the per-hour cost).
4. Click Deploy.

No manual export needed. If you're deploying a raw fine-tune output, the platform automatically builds a serving export behind the scenes — the endpoint deploys when it's ready (a few minutes). You'll see the status move:

Exporting → Verifying → Creating → InService

5. When the endpoint is InService, it's live and serving inference.

Didn't finish? Any deploy stuck in Exporting, Verifying, Creating, or Staged can be deleted from its detail view — teardown stops the in-flight GPU jobs too, so nothing keeps billing.

Step 5 — Call your model (2 min)

From the endpoint detail view, click science Test Inference, type a prompt, and hit Generate — a real response from your owned model.

For programmatic use, the Inference Gateway exposes an OpenAI-compatible API. Generate an API key and call it like any OpenAI endpoint, but pointed at your model:

curl https://api.k3ld.com/api/v1/chat/completions \
  -H "Authorization: Bearer <your-api-key>" \
  -d '{"model":"<your-endpoint-name>","messages":[{"role":"user","content":"Hello!"}]}'

Step 6 — Monitor (1 min)

1. Click Monitor → add Drift schedule.
2. Pick your endpoint and a baseline data path.
3. Set a scan frequency and save.

The platform now scans for drift and flags it before it silently degrades your model's output.

Done 🎉

You have a connected data source, a fine-tuned model you own, a live endpoint, and drift monitoring — with predictable costs the whole way.

Next steps

• Try other presets: SQL generator, Support assistant, or bring your own data.
• Attach a LoRA adapter to a deployed model without retraining from scratch.
• Export your model to GGUF (llama.cpp), ONNX, or MLX and take it anywhere.
• Read the feature guides in features/ or jump to your industry in industries/.

Troubleshooting in one glance