Inference and agents, in the KingdomPay per token · Saudi RiyalDCP-Agent for Saudi business · agents.dcp.saAgents can rent a GPU · npx -y github:dhnpmp-tech/dcp-mcpEarn Riyal from your GPUPDPL · Saudi data residencyInference and agents, in the KingdomPay per token · Saudi RiyalDCP-Agent for Saudi business · agents.dcp.saAgents can rent a GPU · npx -y github:dhnpmp-tech/dcp-mcpEarn Riyal from your GPUPDPL · Saudi data residency
DCP∞
Sign inStart free →
DCP∞
Explore
01Overviewنظرة عامةSovereign Arabic AI runtime→02InferenceالاستدلالOpenAI-compatible · live catalog→03Fine-Tuningالضبط الدقيقLoRA contracts · proof-gated serving→04BatchالدُفعاتJSONL validation · execution gated→05DeploymentsالنشرEndpoint proof · traffic gated→06BenchmarksالقياساتMeasured rows · quality claims gated→07PricingالأسعارPer-million-token · SAR→08GPU Podsحاويات GPURent a whole GPU on demand→09AgentsالوكلاءZero-human onboarding · MCP→10DocsالتوثيقOpenAI-compatible API→11EarnاكسبEarn Riyal from your GPU→12SupportالدعمTalk to the team→
In-Kingdom · PDPL© 2026 · Riyadh
LoRA SFT MVP · proof-gated servingTrain-here/deploy-here contract, honest state

Fine-tuning for Saudi workflows, without pretending the GPU proof is done.

DCP now exposes the LoRA product contract: dataset validation, training-job metadata, model-card stubs, adapter registry rows, deployment intents, endpoint-smoke readiness, usage attribution readiness, and adapter billing readiness. Public managed training, route traffic, adapter usage writes, and adapter billing remain gated until GPU-host artifact proof, vLLM load proof, deterministic endpoint smoke, usage attribution, and money-policy proof exist.

Open Fine-Tuning console ->Read API docs
§ 01 · What is liveContract rails before public claims
dataset_validation

Dataset validation

JSONL format, row counts, token estimate, checksum, and train/validation split are contract-backed.

Statuscontract live
training_job_metadata

Training job rows

Renter-scoped job metadata, model-card stubs, and artifact reservations exist before public trainer execution.

Statuscontract live
adapter_registry

Adapter registry

Adapters can be registered with base model, storage key, checksum, rank, and status; the registry contract proof is now part of the local roadmap suite.

Statuscontract live
artifact_policy

Artifact policy

Adapter and model-card keys are renter/adapter scoped and checksum-guarded before uploads or serving are claimed.

Statuscontract live
deployment_intents

Deployment intents

Deployment rows are visible, and the deployment lifecycle contract proof is part of the local roadmap suite. Route traffic stays off until vLLM proof matches deployment, adapter, endpoint, and artifact.

Statusvisible · routes off
endpoint_smoke

Endpoint smoke readiness

Future endpoint smoke must prove a funded principal, deterministic request, response hash, latency, token totals, and adapter trace.

Statuscontract-only · disabled
usage_attribution

Usage attribution

Future adapter usage rows must carry deployment, adapter, endpoint, checksum, provider, request, token, cost, and pending-settlement fields.

Statuscontract-only · disabled
billing_readiness

Adapter billing readiness

Billing is a disabled policy contract until endpoint smoke, funded-principal, attribution, and settlement proof exist.

Statuscontract-only · disabled
tinker_loop

Tinker-style loop gates

Create LoRA, forward/backward, optimizer step, save, sample, and evaluate are visible as disabled contract gates until GPU proof exists.

Statuscontract-only · disabled
GPU server board used as the visual signal for DCP LoRA fine-tuning and adapter deploymentfig. 02 - LoRA adapter proof path
§ 02 · API surfaceReadiness + deployment intent

Build against the contract now. Serve only after proof.

The public page points developers at the same contract used by the renter console. It is intentionally explicit about what is live, what is metadata-only, and what still needs GPU-host proof.

curl -s https://api.dcp.sa/api/lora/readiness \
  -H "Authorization: Bearer $DCP_RENTER_KEY"

curl -s https://api.dcp.sa/api/lora/readiness \
  -H "Authorization: Bearer $DCP_RENTER_KEY" | jq '.tinker_loop'

curl -s https://api.dcp.sa/api/lora/readiness \
  -H "Authorization: Bearer $DCP_RENTER_KEY" | jq '.adapter_registry.registry_contract_proof'

npm run proof:adapter-registry-contract

curl -s https://api.dcp.sa/api/lora/readiness \
  -H "Authorization: Bearer $DCP_RENTER_KEY" | jq '.adapter_deployments.deployment_contract_proof'

npm run proof:adapter-deployment-contract

curl -s https://api.dcp.sa/api/adapters/artifacts/readiness

curl -s https://api.dcp.sa/api/adapters/endpoints/smoke/readiness

curl -s https://api.dcp.sa/api/adapters/$ADAPTER_ID/deployments/$DEPLOYMENT_ID/endpoint-smoke \
  -H "Authorization: Bearer $DCP_RENTER_KEY"

curl -s https://api.dcp.sa/api/adapters/$ADAPTER_ID/deployments/$DEPLOYMENT_ID/endpoint-smoke \
  -X POST \
  -H "Authorization: Bearer $DCP_RENTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{"funded_smoke_principal":true,"smoke_result":{"request_id":"req_smoke_001"}}'

curl -s https://api.dcp.sa/api/adapters/usage/attribution/readiness

curl -s https://api.dcp.sa/api/adapters/settlement/readiness

curl -s https://api.dcp.sa/api/adapters/billing/approval/readiness

curl -s https://api.dcp.sa/api/adapters/billing/readiness

curl -s "https://api.dcp.sa/api/adapters/deployments?limit=25" \
  -H "Authorization: Bearer $DCP_RENTER_KEY"
  • No public Tinker compatibility claim; low-level loop primitives are contract-only until GPU proof exists.
  • No adapter route traffic until vLLM load proof matches deployment id, adapter id, base model, mode, endpoint id, and checksum.
  • The endpoint-smoke GET returns renter-scoped no-record status, and the POST exists only as a disabled validation contract; it returns 409 and records nothing until response hash, latency, token totals, adapter trace, funded principal, usage attribution, settlement policy, and founder approval pass.
  • No quality claims until reproducible benchmark artifacts exist.
Product boundary

What we are selling now is the path, not an inflated promise.

For today: prepare datasets, inspect training metadata, register adapter artifacts, create proof-gated deployment intents, and inspect disabled endpoint-smoke, usage-attribution, and adapter-billing policy. Next: run LoRA SFT on controlled 3090/4090/5090-class pods, attach artifact proof, load adapters into vLLM, smoke the endpoint with hashed response evidence and adapter trace, and route billed inference only after money and usage evidence exists.

Open consoleRent a GPU pod
01metadata_contract

Readiness, jobs, adapters, and deployment intents are visible now.

02gpu_artifact_proof

Trainer workers must prove the adapter artifact before public training is claimed.

03adapter_load_proof

Serving turns on only when the endpoint proves it loaded the right adapter for the right base model.