KOKO- Local Offline AI Engine & Multimodal AI Agent

A 100% local, privacy-first AI assistant capable of multimodal vision, live system monitoring, and Text-to-SQL database querying. Built with Streamlit, Ollama, and defensive Python engineering.

This project goes beyond basic API wrappers by implementing deterministic intent routing, tool-calling loops, and regex-based output parsing to wrangle Small Language Models (SLMs) and prevent hallucinations.

🧠 Architecture Stack

• Frontend: Streamlit
• LLM Backend: Ollama
• Core Logic Model: llama3.2 (3B Parameters)
• Vision Model: moondream (1.8B Parameters)
• System Tools: psutil, sqlite3, re (Regex)

✨ Key Features

• 100% Offline & Private: Powered entirely by local models running via Ollama. No data ever leaves your machine.
• Multimodal Vision Router: Automatically detects image uploads and routes the prompt to the moondream Vision-Language Model for analysis.
• Deterministic Intent Routing (The Bouncer): Uses zero-latency Python keyword matching to bypass LLM routing hallucinations, securely locking or unlocking system tools based on user intent.
• Live System Monitor: Intercepts hardware queries to execute local psutil commands, feeding real-time CPU, RAM, and Battery data back into the LLM's context window.
• Text-to-SQL Engine (XML Tag Sniper): Bypasses fragile JSON tool-calling schemas by forcing the LLM to write SQL within <SQL> tags. A Python regex sniper extracts the code, runs it against a local SQLite database, and returns the natural language result.