A Platform for Advanced Workflow Creation
For AI-first startups, the challenge isn't just accessing powerful language models — it's orchestrating them into reliable, production-ready workflows that can handle real user interactions. Matter & Gas addresses this challenge with a comprehensive platform that turns AI experimentation into enterprise-grade applications.
The Core Platform: More Than Just Model Access
At its heart, Matter & Gas operates on a graph-based workflow system where AI interactions are built as connected nodes rather than isolated API calls.
Each workflow consists of specialized nodes — from model invocation to document search to conditional routing — connected by edges that define the flow of information.
Nine core node types cover the full spectrum of AI application needs:
- ModelInvoke → AI inference across providers with prompt resolution and token budgets
- Router → conditional branching based on user intent or workflow state
- VectorSearch → retrieval-augmented generation from document collections
- VectorWrite → embedding generation and direct upserts into Pinecone
- SlotTracker → multi-turn slot collection for structured inputs
- ConversationMemory → context persistence across turns
- IntentClassifier → classification-driven routing
- Format → enforce schemas or reshape model output
- StreamToClient → real-time streaming of partial responses with metadata
Workflows maintain a shared GraphState, enabling complex interactions where earlier conversations inform later responses, user intents drive routing, and vector searches provide grounding.
The workflow runner enforces strict safety rules:
- Virtual START/END edges added automatically
- Schema and connectivity validation before execution
- All nodes must be reachable, all router/slot targets must exist
This ensures workflows are deployment-ready before they ever run.
Prompt Management That Scales
A standout feature of Matter & Gas is its centralized prompt engine. Prompts are versioned artifacts, not strings embedded in code:
- BasePromptVersion → immutable prompt content with SHA-256 hash, stored inline or in S3 (KMS-encrypted)
- ActivePromptPointer → mutable selector with rollback history
- CAS Update Lambda → atomic Compare-And-Set prevents race conditions
- PromptArchiveBucket → archival S3 bucket with 1-year retention and daily cleanup of failed uploads
- AuditLog table → every prompt deployment or rollback recorded with author, timestamp, and reason
Resolution hierarchy:
- Tenant + Workflow + Model
- Workflow + Model
- Tenant + Model
- Model (global)
- Emergency fallback → neutral base prompt from
DEFAULT_MODEL_ID
Template Variables
Prompts can interpolate runtime state using ${…} placeholders:
${inputs.user_prompt}→ caller input${workflowId}/${conversationId}→ workflow context${nodeId.output.text}→ prior node outputs${slotTrackerNode.userName}→ collected slot values${context.sessionId}→ session identifier
⚠️ Dot-notation only. No arrays, loops, or conditionals. Missing variables resolve to empty strings with warnings logged.
Multi-Provider AI Integration
Matter & Gas avoids lock-in with a centralized Model Registry. Each model entry defines:
- Identity: provider (OpenAI, Anthropic via Bedrock, Amazon, Meta)
- Capabilities: streaming, JSON mode, function calling, multimodal
- Limits: context windows, reserved output tokens
- Pricing: per 1K tokens, tracked for cost control
- Circuit Breaker: per-model protection with DynamoDB-backed state
Note: Failover between providers (e.g. Claude → GPT-4o) is not automatic.
Instead, workflows select an explicit modelId or rely on the environment's DEFAULT_MODEL_ID.
Circuit Breaker Reliability
Circuit breakers protect apps from cascading failures:
- Trip when error rates exceed thresholds (default: 5% in 5 minutes)
- State persisted in DynamoDB (
ModelCircuitBreaker) - Fail-open if health-check errors occur (prefer availability)
- Manual trip/reset supported with audit logging
Ops teams can force a model offline, then reset when the provider recovers.
Document Intelligence and Vector Search
For knowledge-intensive apps, Matter & Gas provides a document processing pipeline:
- Upload → S3 event triggers async job
- OCR → Amazon Textract processes PDFs/scanned docs
- Chunking & Embedding → Bedrock Titan v2 generates embeddings
- Vector Storage → Pinecone, tenant-scoped namespaces
- Progress Tracking → DynamoDB updates job state
- DLQs → failures routed to
DocumentProcessingDLQorTextractProcessingDLQ
Documents are grouped into Collections:
- Membership controlled via
CollectionMembershipmodel - VectorSearch nodes filter results against
state.allowedDocumentIds - Cascade deletion removes vectors, S3 objects, and DynamoDB records
This transforms static docs into queryable knowledge bases for support, research, or Q&A.
Production-Ready Infrastructure
Matter & Gas runs on AWS Amplify Gen 2 + CDK extensions, hardened for enterprise:
- Observability: Powertools Logger/Metrics/Tracer, CloudWatch metrics, X-Ray traces
- Error Handling: DLQs for all async stages, CAS integrity enforcement, runner timeouts
- Security: KMS encryption, tenant-scoped S3 prefixes, Pinecone namespaces,
WorkflowAccessenforcement - Compliance: PromptArchiveBucket + AuditLog provide immutable history
- Secrets: OpenAI/Pinecone keys pulled from Secrets Manager, never hardcoded
Developer Safety Nets
Developer confidence is built in:
- CAS protection for prompt updates
- Workflow schema validation pre-deployment
- Rollback to prior prompt versions anytime
- Fallbacks everywhere: prompts, circuit breakers, and streaming ensure continuity
Teams can iterate rapidly without breaking production.
Rapid Iteration and Testing
Workflows can be defined in JSON or through the visual builder:
- Simple chatbot → ModelInvoke → Format → StreamToClient
- Complex assistant → IntentClassifier → SlotTracker → VectorSearch → ModelInvoke → Router
Validation ensures correctness, and metrics/logs provide fast feedback.
Real-World Application Patterns
Matter & Gas enables:
- RAG: VectorSearch + ModelInvoke for doc-aware Q&A
- Multi-Turn Conversations: SlotTracker + ConversationMemory
- Intent-Driven Routing: IntentClassifier + Router
- Customer Support: classify queries, search docs, generate answers, escalate when low confidence
- Content Creation: combine user prompts with templates for consistency
The Path Forward
Matter & Gas combines workflow orchestration, prompt versioning, model registry, document intelligence, and production infrastructure into a single foundation.
For AI-first companies, this means:
- Faster time-to-market
- Reliable, compliant applications
- Iteration without fear of regressions
In a rapidly evolving AI landscape, adaptable infrastructure is a competitive edge. Matter & Gas delivers exactly that — enabling teams to build tomorrow's AI applications with the reliability users expect today.