k3ld Model — Customer Documentation
Own your models. Keep your data.
The k3ld Model platform is a guided workspace to connect your data, fine-tune open models, deploy them as your own endpoints, and monitor them — on infrastructure you choose: hosted by k3ld, BYO AWS (your account, k3ld operates it), or self-hosted on AWS (run by you) — with predictable costs, and no per-token vendor bill.
Who this is for
| Reader | Start here |
|---|---|
| First-time user | quickstart.md — live in 15 minutes |
| Analyst / data team | Feature guides in features/ |
| ML / platform engineer | Fine-tuning, post-training, deploy, export, gateway guides |
| Compliance / security reviewer | features/14-security-rbac.md |
| Budget owner | Cost control + Cost optimization |
| Industry-focused reader | End-to-end scenarios in industries/ |
The three stages
Everything in the platform is built around one loop:
Build → Ship → Monitor
- Build — the whole path from raw data to a trained model: connect your data (S3, GCS, warehouses), explore it in a unified catalog with lineage, then create pipelines and train models — guided fine-tuning presets, post-training (DPO/ORPO/GRPO), or CLI. Every run gets a cost estimate and an optional GPU smoke test before it starts.
- Ship — deploy a model as your own real-time endpoint, attach adapters, run canaries, export the weights in any format. Deploying a raw fine-tune builds the serving artifact automatically.
- Monitor — drift scans, health, usage, and self-healing for what you shipped. The loop closes back to Build.
Feature index
| # | Feature | What it does | Guide |
|---|---|---|---|
| 1 | Connect sources | S3 / GCS / warehouse ingestion | 01-connect-sources.md |
| 2 | Explore catalog | Unified catalog, lineage, metastore | 02-explore-catalog.md |
| 3 | Prepare | ETL/ELT pipelines, LLM Format, ML Infer | 03-prepare.md |
| 4 | ML | Train, forecast, classify, cluster — no NN required | 04-ml.md |
| 5 | Train | Launch training runs, CLI | 05-train.md |
| 6 | Fine-tuning | Guided presets (chat, instruction, SQL, support) | 06-fine-tuning.md |
| 7 | Post-training | DPO / ORPO / GRPO preference & reasoning tuning | 07-post-training.md |
| 8 | Ship & deploy | Real-time endpoints, canaries, adapters, auto-export | 08-ship-deploy.md |
| 9 | Export models | safetensors / GGUF / ONNX / MLX / TensorRT | 09-export-models.md |
| 10 | Monitor & drift | Schedules, health, usage, self-healing | 10-monitor-drift.md |
| 11 | Inference gateway | OpenAI-compatible API for your endpoints | 11-inference-gateway.md |
| 12 | Model catalog | Which open models, and when to pick one | 12-model-catalog.md |
| 13 | Cost control | Estimates, GPU smoke, run-confidence cache | 13-cost-control.md |
| 14 | Security & RBAC | Roles, tenants, data residency, compliance | 14-security-rbac.md |
| 15 | MCP integration | Expose your data to AI agents via MCP | 15-mcp-integration.md |
| 16 | Model governance | Which models for which purpose (Settings → Models) | 16-model-governance.md |
| 17 | One-click catalog | Full-weight frontier models, context presets, honest costs | 17-one-click-catalog.md |
Reference
- Cost optimization playbook — cut AI inference + cloud spend, protect gross margin
- Glossary — plain-language terms
- FAQ — the questions reviewers and new users ask
Industry scenarios
Decision guides per industry — which cost, compliance, and operational choices matter at each step, with links to the concrete how-to guides.