Local AI / Agentic Systems

AI Brain

A portable local AI knowledge and engineering workstation built for grounded answers, repository intelligence, research automation, and human-controlled agentic work.

Status
Personal System / Active
Runtime
Windows 11 · SSD-portable · Offline-capable
Core stack
FastAPI · Next.js · PostgreSQL · ChromaDB
Models
Mistral · Phi-3 · Qwen2.5-Coder 7B via Ollama

A local workstation, not a cloud chatbot.

AI Brain brings knowledge, repositories, research sources, security findings, model routing, and controlled engineering actions into one persistent local environment. Its purpose is not generic conversation; it is to support real technical work with inspectable sources and explicit boundaries.

The system is designed to run from a portable SSD on a Windows 11 workstation. When the required local resources are present, it can work offline and does not depend on paid cloud AI APIs.

Portability is an architecture decision.

The SSD-based design keeps the workstation, knowledge, models, and persistent context transportable while preserving a local-first operating model.

Sources are indexed, routed, and grounded locally.

Knowledge and task flow

Knowledge Sources
Ingestion / Indexing
ChromaDB + PostgreSQL
Task / Source Routing
Local Models via Ollama
Grounded Response / Agentic Workflow

Persistent context from varied technical sources.

Local RAG

Indexes PDFs, books, pasted notes, GitHub repositories, documentation sites, trusted crawler research, and security scan findings.

Repository intelligence

Supports local cloning, code and document indexing, repository-specific retrieval, and ChromaDB-backed search.

Research automation

Uses topic classification, source-quality scoring, deduplication, trusted-only mode, approval workflows, and recursive crawling.

Security analysis

Normalizes Semgrep and Bandit findings, then uses local Ollama models to add security explanations.

Automation remains bounded by human approval.

The agentic coding mode can plan changes, inspect repository context, propose edits, preview diffs, and suggest commands. Sensitive actions are deliberately gated.

  • Writes require approval before files change.
  • Commands require explicit approval before execution.
  • Dangerous commands are blocked.
  • Audit logs preserve an inspectable record of agentic activity.
  • Source-grounded context is preferred over unsupported generation.

Different stores serve different responsibilities.

PostgreSQL

Maintains durable application state and persistent chat history.

ChromaDB

Supports vector retrieval for indexed knowledge and repository context.

Local cache

Keeps frequently needed workstation data close to the runtime.

Model routing

Routes work across Mistral, Phi-3, and Qwen2.5-Coder 7B through Ollama where relevant.

A full local engineering environment.

Frontend

Next.js · TypeScript · Tailwind CSS

Backend

FastAPI · SQLAlchemy · Pydantic

Data

PostgreSQL · ChromaDB · Local cache

Local AI

Ollama · Mistral · Phi-3 · Qwen2.5-Coder 7B