FAQ
The questions reviewers and first-time users actually ask. If yours isn't here, it becomes an entry.
Do I need ML experience?
No. The guided UI (presets, cost estimates, GPU smoke, confidence indicators) is designed for non-ML users; the CLI mirrors the same flows for those who want them. See the Quick Start.
Does my data leave my environment?
By default, no migration. The platform reads metadata from your connected sources and there's no export or upload step — you connect the source you already have. Where data is processed: in hosted mode, training/serving executes on k3ld-managed compute (your data is read there during the run); with BYO AWS / self-hosted on AWS, reads and processing stay inside your own account or environment. See 14-security-rbac.md.
Is it per-token billing?
No. Training/fine-tuning and serving are billed on instance time (an hourly rate for the GPU you choose), with a cost estimate before every run. A small credit meter covers platform LLM calls (Ask AI, LLM Format, gateway, storage overage) — see 13-cost-control.md.
Can I take my model with me?
Yes — export is a first-class feature. safetensors, GGUF (llama.cpp/CPU/edge), ONNX, MLX (Apple Silicon), TensorRT-LLM (NVIDIA serving). The weights are your asset. See 09-export-models.md.
What models can I use?
Three surfaces: 22 managed models (DeepSeek, Nova, Llama, Mistral, Qwen, GPT-OSS, Gemma, Claude, Titan) used by platform features; any HuggingFace base you can fine-tune with LoRA/QLoRA; and the one-click frontier catalog (DeepSeek V4 Flash, Qwen3.8-Flash-Next) for full-weight serving. See 12-model-catalog.md.
Who decides which models people can use?
Your admin. Settings → Models controls which models are enabled for which purpose, tenant-wide and per user — the same pattern as the Instances tab for GPUs. See 16-model-governance.md.
How do I know a run won't fail and burn money?
Three rails: a cost estimate before every run, an optional GPU smoke test (validates data + config in minutes, at pennies), and the run-confidence cache (🟢/🟡/🔴 from past identical/near-identical runs). See 13-cost-control.md.
What happens if a deploy fails?
It fails fast with the real reason (a broken artifact is caught at export/verify, before an expensive dead endpoint). Optional self-heal retries automatically; if it escalates, an admin sees the real error with one-click approve/stop. Any deploy — even stuck ones — can be deleted; teardown stops in-flight GPU jobs so nothing keeps billing. See 08-ship-deploy.md.
I'm done with an endpoint — what now?
Delete it, or use keep-for-later (≈ $0 + small S3 storage, one-click resume). Deployed endpoints bill by uptime — don't leave idle ones running. See 13-cost-control.md.
Can I use my own AWS account?
Yes — three options. Hosted (default): k3ld runs everything on its managed infrastructure — no AWS account needed, most frictionless. BYO AWS: you bring the account and k3ld deploys/operates into it. Self-hosted on AWS: the full platform runs in your environment, operated by you. BYO and self-hosted inherit your existing security approvals and procurement. See 08-ship-deploy.md.
Is the platform SOC 2 / ISO certified?
Not claimed here. The platform enforces industry-standard security controls (MFA, strong password policies, brute-force protection, encrypted transport, tenant isolation, and an audit trail) — see 14-security-rbac.md — but has not been independently certified as of this writing. Ask the vendor team where they are on their certification roadmap, or for the full control details under NDA, if you need attestation.
Can my AI agents reach my data?
Yes, via MCP. The platform exposes your governed catalog/data/models as MCP tools, scoped by the same RBAC — agents inherit your permissions, not a master key. See 15-mcp-integration.md.
How do I call my model from code?
The Inference Gateway is an OpenAI-compatible API (https://api.k3ld.com/api/v1/...) — chat completions, embeddings, rerank. Existing OpenAI-style code works with a URL + key + model swap. See 11-inference-gateway.md.