AegisMind is a modular, agentic AI system built to deeply understand and demonstrate how modern multi-agent AI architectures work in practice. The project focuses on orchestration, memory, retrieval, and tool-based reasoning rather than relying on monolithic LLM behavior.
It uses the Groq API purely as a fast inference engine, while all intelligence — planning, routing, memory, document understanding, and tool usage — is engineered explicitly.
Goal: Learn and showcase true Agentic AI system design, not just prompt engineering.
-
Multi-Agent Architecture
- Planner agent for intent detection and routing
- Specialized agents for document Q&A, memory, voice, and tools
-
Agent Orchestration (LangGraph)
- Explicit state management
- Conditional routing between agents
- Clear separation of planning vs execution
-
Document Intelligence (RAG)
- Upload PDFs and documents
- Chunking and embeddings
- Retrieval-augmented generation for grounded answers
-
Memory System
- Short-term conversational memory
- Long-term semantic memory (facts, preferences, summaries)
- Memory write/read rules to avoid context pollution
-
Groq-powered Reasoning
- Uses Groq-hosted LLaMA / Mixtral models
- Stateless inference with structured prompts
-
Voice Support (Optional Module)
- Speech-to-text (Whisper)
- Text-to-speech (Piper / Coqui)
-
MCP (Model Context Protocol) Support (Advanced)
- MCP client integration
- Agent-based tool invocation via MCP servers
Most chatbot projects rely on:
- Single-agent prompts
- Implicit memory (chat history)
- Hard-coded tool calls
AegisMind is different:
- LLMs are treated as stateless reasoning engines
- Intelligence emerges from system design, not prompts
- Memory, planning, and tools are explicit, inspectable components
This mirrors how real-world AI agent systems are built.
User (Text / Voice)
↓
Orchestrator (LangGraph)
↓
┌─────────────────────────────┐
│ Planner Agent │
│ ─ Intent & routing │
└──────────────┬──────────────┘
↓
┌──────────────────────────────────────────┐
│ Execution Agents │
│ - Document QA (RAG) │
│ - Memory Agent │
│ - MCP Tool Agent │
│ - Voice Agent │
└──────────────┬───────────────────────────┘
↓
Groq LLM API (Reasoning Only)
aegismind/
│
├── agents/
│ ├── planner.py # Intent detection & routing
│ ├── doc_qa.py # RAG-based document agent
│ ├── memory.py # Memory read/write logic
│ ├── voice.py # STT / TTS handling
│ └── mcp_agent.py # MCP tool integration
│
├── orchestration/
│ ├── graph.py # LangGraph definition
│ └── state.py # Global agent state schema
│
├── rag/
│ ├── loader.py # Document loading
│ ├── embedder.py # Embedding generation
│ └── retriever.py # Vector search
│
├── memory/
│ ├── short_term.py # Session memory
│ └── long_term.db # Persistent memory store
│
├── services/
│ └── groq_client.py # Groq API wrapper
│
├── ui/
│ ├── cli.py # CLI interface
│ └── web.py # (Optional) Web UI
│
└── main.py # Entry point
- User sends text or voice input
- Planner agent classifies intent
- Orchestrator routes request
- Relevant agents execute tasks
- Groq LLM performs reasoning
- Memory is updated (if needed)
- Final response returned (text/voice)
- LLM Inference: Groq API (LLaMA 3 / Mixtral)
- Agent Orchestration: LangGraph
- Embeddings: HuggingFace (CPU-friendly models)
- Vector Store: FAISS / Chroma (local)
- Memory Storage: SQLite / JSON
- Voice: Whisper (STT), Piper / Coqui (TTS)
- Protocol: MCP (Model Context Protocol)
- Day 1–2: Core architecture & orchestrator
- Day 3–4: Multi-agent routing
- Day 5–6: Document upload & RAG
- Day 7: Memory system
- Day 8: Memory + agents integration
- Day 9: Voice or MCP integration
- Day 10: Cleanup, README, demo
By completing this project, you will understand:
- How real agentic AI systems are structured
- Why planning must be separated from execution
- How to design reliable memory for LLMs
- How to ground LLM responses using RAG
- How to integrate tools via open protocols (MCP)
This project prioritizes learning and architectural correctness over production scale. It is designed as a flagship portfolio project for AI/ML and Agentic AI roles.
MIT License