AI/ML Platform

Agentic AI infrastructure powering Norns and related services.

Overview

The AI/ML platform provides intelligent automation through LangGraph-based agents, local LLM inference, and comprehensive observability.

┌──────────────────────────────────────────────────────────────────┐
│                        AI/ML PLATFORM                             │
├──────────────────────────────────────────────────────────────────┤
│                                                                   │
│   ┌─────────────────┐    ┌─────────────────┐    ┌─────────────┐  │
│   │   Norns Admin   │───▶│   LangGraph     │───▶│   Ollama    │  │
│   │   (Frontend)    │    │   (Agents)      │    │   (LLM)     │  │
│   └─────────────────┘    └─────────────────┘    └─────────────┘  │
│            │                     │                     │          │
│            ▼                     ▼                     │          │
│   ┌─────────────────┐    ┌─────────────────┐          │          │
│   │  Norns Agent    │───▶│   Langfuse      │◀─────────┘          │
│   │   (Backend)     │    │ (Observability) │                     │
│   └─────────────────┘    └─────────────────┘                     │
│            │                                                      │
│            ▼                                                      │
│   ┌─────────────────┐                                            │
│   │   Redpanda      │                                            │
│   │  (Messaging)    │                                            │
│   └─────────────────┘                                            │
└──────────────────────────────────────────────────────────────────┘

Services

Service	Purpose	URL
Norns Agent	Core AI agent backend	norns-pm.ravenhelm.dev
Norns Memory System	Long-term memory & learning	(internal)
Norns Admin	Admin UI & chat interface	norns.ravenhelm.dev
Norns Web Crawling	Web content ingestion for RAG	norns.ravenhelm.dev/documents
LangGraph	Agent workflow engine	norns-api.ravenhelm.dev
Langfuse	LLM observability	langfuse.ravenhelm.dev
Ollama	Local LLM inference	Internal (11434)
Redpanda	Event streaming	Internal (9092)
User Profile System	Profile management & insights	bifrost-api.ravenhelm.dev
Bifrost MCP Gateway	Tool routing & execution	bifrost-api.ravenhelm.dev
Bifrost Tool Catalog	Complete tool reference (21 tools)	-
Norns Telephony	Outbound/inbound calling	+1 (737) 214-3330
Norns WhatsApp Channel	WhatsApp messaging	Twilio Sandbox
Multi-Channel Sessions	Cross-channel conversation memory	-

Architecture

Agent Flow

User Request → Norns Admin → LangGraph → Agent Graph
                                            │
                                            ├── Memory (Muninn - episodic/semantic)
                                            ├── Tools (Home Assistant, Calendar, etc.)
                                            ├── Ollama (embeddings, local inference)
                                            └── Claude API (complex reasoning)
                                            │
                                            ▼
                                         Langfuse (trace recording)

Memory Architecture (Raven Model)

┌─────────────────────────────────────────┐
│ Huginn (State) - Redis session state    │
└─────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────┐
│ Context (Identity) - User permissions   │
└─────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────┐
│ Muninn (Memory) - Episodic + Semantic   │
│  • Vector embeddings (pgvector)         │
│  • Pattern extraction & learning        │
│  • Knowledge entity graph               │
└─────────────────────────────────────────┘

See Norns Memory System for details.

Data Stores

PostgreSQL (ravenmaskos): Tasks, projects, user data, memories
pgvector: Embedding storage for semantic search
Redis: Session cache, rate limiting
Redpanda: Event streaming between agents

Capabilities

Norns Agent Capabilities

Capability	Description
Task Management	Create, update, query tasks
Calendar	Google Calendar integration
Home Control	Home Assistant device control
Grocery	Grocy inventory management
Observability	Query Grafana, Loki, Prometheus
Slack	Post messages, respond to commands
Memory	Recall past interactions, learn patterns
Knowledge	Web search, weather, news (details)
Telephony	Make and receive phone calls (details)

Agent Domains

Norns uses a hierarchical agent architecture with specialized domain supervisors:

Domain	Description	Documentation
Task	Task and project management	Norns Agent
Calendar	Google Calendar integration	Norns Agent
Home	Smart home control via Home Assistant	Norns Agent
Navigation	Directions, places search, travel times	Navigation Domain
Observability	Metrics, logs, and monitoring	Norns Agent
Memory	Long-term memory and learning	Norns Memory System
Knowledge	Web search, weather, news	Knowledge Domain
Code	File operations, git, GitLab, CI/CD	Coding Agent
Social	Slack messages, DMs, notifications	Norns Agent
Goal	Goal tracking and background execution	Goals & Background Execution

Quick Commands

# Check all AI/ML services
for svc in norns-agent norns-admin langgraph langfuse-web ollama norns-redpanda; do
  echo "=== $svc ==="
  docker ps --filter "name=$svc" --format "{{.Names}}: {{.Status}}"
done

# View Norns agent logs
docker logs -f norns-agent

# Test Ollama
docker exec ollama ollama list

# Query Langfuse
curl https://langfuse.ravenhelm.dev/api/public/health

# Check memory stats
docker exec -i postgres psql -U ravenhelm -d ravenmaskos -c \
  "SELECT 
    (SELECT COUNT(*) FROM episodic_memories) as episodic,
    (SELECT COUNT(*) FROM semantic_patterns) as patterns;"

Environment Variables

Key environment variables for AI services:

# API Keys
ANTHROPIC_API_KEY=<key>
OPENAI_API_KEY=<key>

# Langfuse
LANGFUSE_PUBLIC_KEY=<key>
LANGFUSE_SECRET_KEY=<key>
LANGFUSE_HOST=http://langfuse-web:3000

# Embeddings
OLLAMA_URL=http://ollama:11434
OLLAMA_EMBED_MODEL=nomic-embed-text
EMBEDDING_PROVIDER=auto

# Integrations
SLACK_BOT_TOKEN=<token>
HOMEASSISTANT_TOKEN=<token>
GOOGLE_CALENDAR_CLIENT_ID=<id>

Recent Updates

Fixed and documented the Navigation domain supervisor with several critical improvements:

Two-phase execution: Automatically gets directions to top result when user requests "find X and get directions"
Database-first location resolution: Prioritizes users.location over memory context for more reliable "near me" queries
Slack delivery integration: Navigation handles its own Slack delivery (only sends directions, not places list)
Parallel routing fix: Prevents Social domain from being called in parallel with Navigation (avoiding duplicate messages)
Graceful error handling: Converts all API errors to user-friendly messages

See Navigation Domain for complete documentation.

RBAC Integration (2026-01-03)

Norns Admin and Bifrost API now integrate with OpenFGA for fine-grained authorization:

Store ID: 01KE1W1JS5HZWWJSZKG5HR9XA5
Model ID: 01KE1W3RJH1E13G84N3ERN5XDN
Tuples migrated: 23 authorization relationships
Integration: Bifrost's services/openfga.py and auth/dependencies.py
CORS workaround: Norns Admin uses /api/rbac/permissions proxy

See OpenFGA for details.

Memory System Fix (2026-01-03)

Fixed pgvector embedding handling in Norns memory system:

Issue: Embeddings were incorrectly converted to strings
Fix: Pass embeddings as Python lists directly to asyncpg
Affected files: episodic.py, muninn_context.py
Type registration: pgvector.asyncpg.register_vector() handles conversion

See Norns Memory System for details.

Zitadel Opaque Token Support (2026-01-03)

Bifrost now handles both JWT and opaque tokens from Zitadel:

Detection: token.count('.') == 2 → JWT, else opaque
Validation: Opaque tokens validated via Zitadel's /oidc/v1/userinfo
User lookup: By auth_provider_id (Zitadel sub), not UUID

See Zitadel for token details.

Major update to Norns agent with parallel task execution and communication capabilities:

Multi-Task Support:

Compound requests are now automatically decomposed into discrete tasks
Tasks are executed in parallel using asyncio.as_completed() for optimal performance
Each task streams its own progress independently
Example: "Add eggs to shopping list, check calendar, and call John" → 3 parallel tasks

Social Domain:

New SOCIAL domain for all communication tasks
SlackMessagingWorker: Send Slack messages and DMs
PhoneCallWorker: Make outbound calls via telephony service (with API key auth)
SMSWorker: Send SMS via Twilio REST API

Smart Contextual Greeting:

Greetings now vary based on conversation context
Uses dynamic templates with time-of-day awareness
Fresh starts get friendly greetings; quick follow-ups get "Sure!" style responses

Telephony Delegation:

Voice agent can now delegate to Norns for Slack messaging during calls
Updated system prompt to allow broader delegation types
User can say "Send me that on Slack" and the voice agent will delegate

See Norns Agent for full documentation.

Overview​

Services​

Architecture​

Agent Flow​

Memory Architecture (Raven Model)​

Data Stores​

Capabilities​

Norns Agent Capabilities​

Agent Domains​

Quick Commands​

Environment Variables​

Recent Updates​

Navigation Domain Supervisor (2026-01-05)​

RBAC Integration (2026-01-03)​

Memory System Fix (2026-01-03)​

Zitadel Opaque Token Support (2026-01-03)​

Multi-Task Parallel Execution & Social Domain (2026-01-07)​