Local AI Platform

Full-stack AI platform: FastAPI backend + Flutter desktop/web UI + local LLM via Ollama.

Target hardware: RTX 4060 Laptop (8GB VRAM), 32GB RAM, Windows 11.

Features

Chat & Agents

• Multi-model chat (Ollama, HuggingFace, LM Studio) with SQLite persistence
• LangGraph agent system with tools, MCP servers, and multi-agent DAG orchestration
• Streaming SSE, conversation history, file attachments

Image Generation

• Stable Diffusion pipelines (SD 1.5, SDXL) with TF32/SageAttention optimizations
• Prompt enhancement via local LLM
• Model catalog with auto-detection

Image Editor (59+ operations)

Classical (CPU, instant): brightness, contrast, saturation, vibrance, clarity, hue, gamma, color temperature, shadows/highlights, auto levels, auto white balance (learning-based), HDR tone mapping (Reinhard/Drago/ACES/Mantiuk), tiered denoising (NLMeans/BM3D/wavelet), skin smoothing (guided filter), LAB-space sharpening, dehaze with sky protection, 5 built-in 3D LUTs (.cube), 6 presets (research-backed), lens corrections, morphological ops, FFT filtering, vignette, grain, and more.

AI (GPU): FLUX.1-Kontext-dev instruction editing (GGUF-quantized, cpu_offloaded for 8GB cards — see KONTEXT_PIPELINE.md), CosXL Edit, InstructPix2Pix (DPMSolver, 20-step), face restoration (GFPGAN + CodeFormer), super-resolution (RealESRGAN), background removal (BiRefNet/U2-Net), portrait bokeh (Depth Anything v2 ONNX), neural style transfer (5 ONNX styles), LaMa inpainting with brush UI, algorithmic colorization, low-light enhancement, smart auto-enhance with BRISQUE scoring.

Non-destructive undo/redo history, session persistence, PNG/JPEG/WEBP export.

AI Partner

• Conversational companion with Big Five personality profiling
• Emotional voice (Chatterbox TTS, separate venv)
• Mem0 + ChromaDB memory, wellness detection, anti-manipulation safeguards

Quick Start

Backend

python -m venv .venv && .venv\Scripts\activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install -e .
pip install -e ".[editor]"      # Optional: image editor AI features
pip install -e ".[partner]"     # Optional: AI partner
uvicorn api_server:app --host 127.0.0.1 --port 8000 --reload

Flutter Client

cd flutter_client && flutter pub get && flutter run -d windows

Chatterbox TTS (optional)

python -m venv .venv_chatterbox && .venv_chatterbox\Scripts\activate
pip install chatterbox-tts && python scripts/chatterbox_server.py

Data Directories

All directories auto-created. Model weights download on demand.

Configuration

.env file in project root:

# Core
OLLAMA_DEFAULT_MODEL=gemma3:4b
HF_TOKEN=hf_...                      # required for gated models (FLUX Kontext, FLUX.1-dev)
TAVILY_API_KEY=tvly-...

# FLUX Kontext image editor (see KONTEXT_PIPELINE.md for full details)
KONTEXT_GGUF_QUANT=Q3_K_S            # Q2_K | Q3_K_S (recommended for 8GB) | Q4_K_S | Q4_K_M | Q5_K_S | Q6_K | Q8_0
# KONTEXT_FBC_THRESHOLD=0.08         # leave unset to DISABLE FirstBlockCache (required for strong edits)
# DIFFUSERS_GGUF_CUDA_KERNELS=true   # Linux only — breaks diffusers import on Windows

Architecture

api_server.py                       # FastAPI (all endpoints)
src/local_ai_platform/
  agents.py                         # LangGraph orchestrator
  config.py / db.py                 # Config + SQLite
  providers/                        # Ollama, HuggingFace, LM Studio
  images/
    processors.py                   # 48 classical operations
    ai_enhance.py                   # AI models (IP2P, GFPGAN, RealESRGAN, etc.)
    ai_models.py                    # ONNX models (style, inpaint, colorize, depth)
    editor.py                       # Editor service (sessions, undo/redo)
    service.py                      # Image generation service
  partner/                          # AI companion (engine, profile, memory)
  tools/                            # Agent tools (code exec, search, etc.)
flutter_client/lib/pages/           # Flutter UI
scripts/chatterbox_server.py        # Standalone TTS server

Documentation

Long-form reference docs live at the repo root (same place as this README) so you don't have to hunt through subdirectories:

If you hit a problem with Kontext specifically — weak edits, slow steps, device mismatch errors, OOM, cache confusion, quant selection — read KONTEXT_PIPELINE.md first. Most known issues are documented there with their root cause and the fix.

License

MIT