5 AI RUNTIMES

Every AI execution model.
One platform.

From LLM gateway to GPU containers to agent loops — pick the runtime your app needs. Each one is production-ready, not a roadmap item.

5
Runtimes
3
Deploy environments
Real
GPU containers
Zero
Shared infra
BRICQSruntimeLLM APIGPUAgentRAGWorkflow
GATEWAY

LLM API Gateway

Connect OpenAI, Anthropic, or Google Gemini behind one unified OpenAI-compatible endpoint. Switch providers without touching app code. All provider keys stored in your org's dedicated secrets vault — never in the database.

  • OpenAI-compatible /v1/chat/completions
  • Provider-agnostic routing
  • Keys encrypted in org vault
  • Usage tracked per org
  • Fallback routing on provider error
OpenAI-compatibleAnthropicGoogle GeminiKey vaultUsage tracking
# One endpoint, any provider
curl https://api.bricqsai.com/llm/v1/chat/completions \
-H "Authorization: Bearer $BRICQS_KEY" \
-H "X-Provider: anthropic" \
-d '{
"model": "claude-sonnet-4-5",
"messages": [{"role": "user", "content": "Hello!"}]
}'
# Switch to openai — zero code change
-H "X-Provider: openai"
DEVLLaMA 3 8BGPU: A10promotePREVIEWLLaMA 3 8BGPU: A10promoteLIVEPRODLLaMA 3 8BGPU: A100CPU 12%MEM 34%GPU 71%Requests/s 142P99 88ms
CONTAINERS

GPU Container Runtime

Deploy open-weight models — LLaMA 3, Mistral, Phi-3, Qwen, Gemma — to dedicated GPU containers on Azure. Three environments (development, preview, production) with zero-downtime promote. Real CPU/memory/GPU% metrics from Azure Monitor, no instrumentation needed.

  • LLaMA 3 8B / 70B, Mistral 7B, Phi-3, Qwen 2, Gemma 2B
  • Dev / preview / prod environments
  • Zero-downtime production promote
  • Real Azure Monitor metrics (CPU% / mem% / GPU%)
  • Container logs streaming
LLaMA 3MistralPhi-3Qwen 2Azure MonitorGPU
TOOL USE

Agent Runtime

A structured reasoning loop with sandboxed tool access — calculator, web fetch, and code execution. Up to 6 reasoning steps, full trace returned. Powered by Anthropic Claude for the reasoning backbone.

  • Sandboxed tool execution
  • Web fetch + calculator tools
  • Full step trace returned
  • Up to 6 reasoning steps
  • Anthropic Claude backbone
Tool useSandboxedStep traceClaude backboneWeb fetch
Step 1ReasonStep 2web_fetchStep 3ReasonResultAnswerTrace→ [1] Reasoning: "I need to fetch the current gold price..."→ [2] Tool: web_fetch("https://metals-api.com/gold") → $2,318/oz→ [3] Reasoning: "Price confirmed. Calculating portfolio delta..."✓ Result: Portfolio up $4,820 (2.1%) based on current spot price.
PDFQ4_Report↑ uploadedChunker512-token+ embedVector IndexQueryQ: "What was the gross margin in Q4?"→ 3 chunks retrieved (cos sim > 0.82)A: "Gross margin was 67.3% in Q4 [source: Q4_Report.pdf, p.8]"
RETRIEVAL

RAG — Retrieval-Augmented Generation

Upload PDF, DOCX, or TXT documents. BRICQS chunks, embeds, and indexes them automatically. Ask questions and get answers grounded in your documents via vector similarity search.

  • PDF / DOCX / TXT upload
  • Automatic chunking + embedding
  • Vector similarity search
  • Grounded answers with source citations
  • Document management UI
PDFDOCXTXTVector searchCitations
PIPELINES

Workflow Pipelines

Chain prompt-template steps into a multi-step AI pipeline. Each step can call any other runtime — a GPU deployment, a RAG index, or the LLM API. Pass outputs forward with {{previous}} or reference named steps.

  • {{input}} / {{previous}} templating
  • Cross-runtime step composition
  • Step-by-step execution log
  • Visual canvas builder
  • Independent deployment with its own endpoint
Multi-stepCross-runtimeTemplate varsVisual canvasAuto deploy
Step 1ClassifyLLM APIStep 2RetrieveRAGStep 3GenerateGPUOutputReturnJSONTemplate varsStep 1 prompt: "Classify this request: {{input}}"Step 2 prompt: "Find docs relevant to: {{step_1.output}}"Step 3 prompt: "Using {{previous}}, write a detailed answer."

Pick the right runtime

Each runtime is independent. You can use one or all five within the same project.

RuntimeUse when…InputOutputGPU?
LLM APIYou need one API key for all providersMessages arrayChat completion—
GPU RuntimeYou need private, open-weight modelsMessages / promptCompletion✓
AgentYou need multi-step reasoning + toolsGoal promptResult + trace—
RAGYour app needs document Q&ADocuments + questionGrounded answer—
WorkflowYou need a reusable AI pipelineVariablesFinal step outputOptional