# What is k3ld? — Overview Q&A

Grounded answers from verified platform behavior. (Source: k3ld site bot — post-training
multi-turn conversations, seeded from real, verified platform behavior.)

## What is k3ld?

k3ld is an AI platform where you own the whole stack — your data, your models, and your endpoints. There are two main consoles: **platform.k3ld.com** for data work (catalog, pipelines, analytics, ML) and **model.k3ld.com** for building and shipping your own models, plus **mcp.k3ld.com** for AI-agent tooling. The tagline is "Own Your Intelligence" — no black-box vendor lock-in.

## Who is it for?

Three groups:

- **Data teams** that want a catalog and pipelines without standing up Spark clusters;
- **ML engineers** who want to train, fine-tune, and deploy models in one place;
- **Businesses** that need own-your-data AI.

One sign-on gets you both consoles.

## How is it different from Snowflake + Databricks + Hugging Face separately?

It's one journey instead of three tools bolted together. Explore → Build → Train → Ship → Monitor all happen in one place, on one credit meter, with the same RBAC and audit trail. Concretely: a single pipeline can read a catalog table, redact PII, train a model, and deploy an endpoint — and the platform routes the compute for you. SQL work runs serverless, heavy ML training runs on EMR, and neural-net training runs on SageMaker. You never touch a cluster console.

## What does "Own Your Intelligence" actually mean?

The artifacts are yours: your catalog tables, your trained model weights, your deployed endpoints, your data capture logs. You can download or export your models at any time. If you leave, you take the intelligence with you — that's the difference from renting a model from a black box.

## Raw data to a deployed model — fastest path

Upload to your Managed S3 bucket (every account gets one, no connection config needed). In platform.k3ld, build a pipeline: File Source → PreProcessor (auto) → Train Model → Evaluator → Reporter → S3 Export. Pick the target column and hit Run. The pipeline runs serverless — no cluster to manage — and Train Model saves the model artifact.

Use a SchemaScout node first to see column types, cardinality, and null percentages; PreProcessor auto-handles impute/encode/scale. The editor has roughly **80 node types** across inputs, transforms, ML, flow control, and NN — you can branch, merge, and loop.

Use the model three ways: a **Classify node** inside a pipeline, an **ML Infer node** that scores rows in place, or — most common — **deploy it as an endpoint** on model.k3ld. Endpoints auto-scale, auto-sleep when idle to stop billing, and capture data so you can audit predictions. Schedule pipelines with cron, trigger by webhook, or use a Gate node.

## Pricing

Everything runs on **credits — 1 credit ≈ $0.01**. Pipelines, notebooks, embeddings, model training, and endpoint uptime all meter to the same balance; a usage panel breaks it down by feature. Credits reset on the 1st of each month.

It's transparent before you commit: before launching a pipeline or training job, the platform shows an estimate — credits, hours, and per-minute rate per instance, with spot vs on-demand options. The final charge reflects the actual instance and uptime.

### Tiers

| Tier | Credits/month | Compute | Notes |
|---|---|---|---|
| Free | exploring | serverless CPU + EMR, GPU up to a modest instance | start here |
| Pro | 50,000 | serverless CPU + EMR, GPU up to g5.48xlarge | |
| Scale | 250,000 | p4d-class GPUs, AI agent | |
| Enterprise | 500,000 | all compute up to H200 (ml.p5e.48xlarge), row-level security + masking, 99.9% SLA, 4-hour support | |
| Enterprise+ | negotiated | unlimited credits, dedicated EMR cluster, VPC deployment | |

If a run needs an instance above your tier, the platform tells you before you spend, and you can upgrade in place.

**Spot:** ~65% cheaper; the platform tries spot first and falls back to on-demand if capacity is unavailable. Great for batch jobs and training that can tolerate interruption.

## Security & privacy

- Every tenant is fully isolated — separate metastore, separate buckets, separate compute — and data is encrypted at rest in S3.
- **PII Detect** and **PHI Detect** nodes (PHI uses AWS Comprehend Medical, HIPAA-eligible); choose to redact, replace, or flag findings, running serverless.
- Roles (admin, editor, viewer) control access; every sensitive action lands in an audit log with 90-day retention.
- Deployed model endpoints are private to your tenant by default; you expose them deliberately.