Memory Bus (Event System)
An MQTT-style in-memory event bus that coordinates every component without coupling them together. Topic-based publish/subscribe with wildcard matching, event history replay, backpressure handling, and thread-safe operation across async and sync contexts.
Pub/Sub Wildcards
Event Replay
Backpressure
Thread-Safe
Why it matters: Traditional event systems force you to choose between async and sync. Ours handles both in the same bus — with MQTT-style hierarchy for granular routing. agent/# catches everything an agent does. llm/* catches only direct LLM events.
Semantic Cache
A three-tier intelligent caching layer that sits between your application and any LLM provider. It doesn't just match exact queries — it understands when two questions mean the same thing and serves cached responses, cutting LLM API costs by 30–50% in production.
Exact Hit
Context Hit
Tag Filtering
TTL Support
Why it matters: "What is machine learning?" and "Explain ML to me" are different strings but the same question. Our semantic cache recognises this, serving the cached response instead of burning another API call. At scale, this saves thousands.
FAISS Vector Store
A multi-strategy vector search engine supporting three index types optimised for different scales — from exact nearest-neighbour for precision to HNSW graph indexing for large-scale sub-millisecond retrieval. GPU-accelerated with automatic CPU fallback.
Flat / IVF / HNSW
GPU Accelerated
Async Operations
Persistent Storage
Why it matters: Different workloads need different search strategies. Small dataset needing perfect recall? Use Flat. Millions of embeddings? HNSW gives you sub-millisecond lookups. The system picks the right strategy — or lets you choose.
Provider-Agnostic LLM Layer
A unified abstraction that lets your application talk to any LLM backend — Ollama, Claude, OpenAI, vLLM, llama.cpp — through a single interface. Hot-swap providers at runtime through config changes alone, with zero code modifications.
Hot-Swap Providers
Streaming
Tool Calling
Health Checks
Why it matters: Locked into one AI provider? That's a risk. Our abstraction means you can run Ollama locally during development, switch to Claude in production, and fall back to OpenAI if needed — all from a config file. No vendor lock-in.
Multi-Agent Orchestration
A complete agent lifecycle manager using the ReAct (Reasoning + Acting) pattern. Agents are defined in YAML, spawned dynamically, and coordinated through tool execution, state tracking, and parent-child relationships — with token budgets and step limits built in.
ReAct Pattern
YAML Definitions
State Machine
Tool Coordination
Why it matters: Single-agent systems hit a ceiling fast. Our orchestrator manages full agent lifecycles — spawning, reasoning, tool use, completion — with guardrails (token limits, step caps) that prevent runaway costs or infinite loops.
8-Dimensional Health Scoring
A composite health intelligence engine that fuses cardiac recovery, sleep quality, neuro-recovery, fatigue index, injury risk, cognitive readiness, pain prediction, and fat adaptation into real-time personalised scores with context-aware alerts.
Cardiac Recovery
Neuro-Recovery
Injury Risk
Cognitive Readiness
Why it matters: Most health platforms track one or two dimensions. Our engine weights 8 composite metrics from biometric, environmental, and behavioural data — then generates severity-graded alerts with personalised recommendations in real time.
Orchestrator