🤖 MyLLM – Personal AI Stack on Your Mac

Run a complete private AI ecosystem locally using Ollama, Open WebUI, and purpose-built tools for email, finance, 3D generation, image generation, speech-to-text, and social media marketing. No cloud. No subscriptions. No data leaving your machine.

🧰 Tools in This Stack

Tool	Purpose	Port
💬 Open WebUI + SearXNG	Chat UI with private web search	:3000 / :8080
📈 Quant AI	Stock & crypto analysis + portfolio tracker	:8000
📧 Mail AI Manager	AI email classifier for any IMAP provider (Gmail, iCloud, Outlook, ProtonMail) — 100% local, no Gmail API	:5051
🎨 Stable Diffusion	Local text-to-image generation · Apple Silicon MPS	:5050
🧊 TripoSR 3D Pipeline	Image → 3D mesh (Apple Silicon)	:5050
🤖 Tax AI Social	AI social media content engine for tax/accounting firms	:5055
🎙️ Whisper STT	100% local speech-to-text for Open WebUI	:9000
🧠 AI Model Training Studio	Fine-tune and train AI models on custom data. LoRA training, data processing, export to Ollama.	:8501

🏗️ Architecture

                    🦙 Ollama (localhost:11434)
                    mistral:7b · deepseek-coder:6.7b
                    Apple Silicon MPS · ~4-10GB RAM
                           │
        ┌──────────────────┼──────────────────────┐
        │                  │                      │
        ▼                  ▼                      ▼
  💬 Open WebUI      📈 Quant AI           📧 Mail AI
  :3000              :8000                 :5051
  + SearXNG          FastAPI+ChromaDB      IMAP + LLM
  :8080              vectorbt              Auto-triage

        │                  │                      │
        ▼                  ▼                      ▼
  🧊 TripoSR         🤖 Tax AI Social   🧠 AI Training
  :5053              :5055              :8501
  Image→3D           Instagram/         LoRA/Fine-tune
  OBJ/GLB            Facebook/TikTok    Ollama Export
  Apple Silicon      Compliance         Data Processing

📋 Requirements

macOS (Apple Silicon M1/M2/M3/M4 recommended) or Linux
16 GB RAM recommended (8GB minimum for basic use)
Docker Desktop — https://www.docker.com/products/docker-desktop/
15+ GB free disk space (for models + tools)
Internet connection (for initial setup only)

🚀 Quick Start

Step 1 — Clone the repo

git clone https://github.com/jup313/MYLLM.git
cd MYLLM

Step 2 — Run setup

chmod +x setup.sh install-tools.sh
./setup.sh

This will:

Install Ollama
Download mistral:7b and deepseek-coder:6.7b models
Start Open WebUI on http://localhost:3000
Start SearXNG on http://localhost:8080

Step 3 — Create your account

Open http://localhost:3000
Click Sign Up and create your admin account
The first account is automatically admin

💬 Tool 1 — Open WebUI + SearXNG

Chat with local AI models + private web search.

URL	Description
http://localhost:3000	Chat interface
http://localhost:8080	SearXNG search engine

Enable web search

Click the ✨ sparkle icon in chat input
Toggle Web Search ON
Ask any question — AI searches privately via SearXNG

Install Shopping Search Tool

./install-tools.sh your@email.com yourpassword

📈 Tool 2 — Quant AI

AI-powered stock and crypto analysis running 100% locally.

cd quant_api
pip install -r requirements.txt
python main.py

FastAPI + ChromaDB vector database
vectorbt for backtesting
curl-cffi for market data
Ask natural language questions about stocks and crypto

📧 Tool 3 — Mail AI Manager ⭐ UPDATED

🎉 Now works with ANY IMAP email provider (Gmail, iCloud, Outlook, ProtonMail) — no Gmail API restrictions!

Local AI that classifies, summarizes, categorizes, and drafts replies to your emails.

cd mail-ai-manager
python3 app.py
# Open: http://localhost:5051

What's New in Phase 4 ✨

Migration from Gmail Manager to universal Mail AI Manager:

✅ Works with any IMAP provider — Gmail, iCloud, Outlook, ProtonMail, etc.
✅ No OAuth complexity — Just IMAP credentials (app-specific password)
✅ 100% local — Zero cloud dependencies, zero API restrictions
✅ Hybrid IMAP + AppleScript — Reliable with native macOS Mail.app fallback
✅ Feature parity — All Gmail manager features ported to Mail system
✅ Same AI classification — Work, spam, personal, urgent categorization
✅ Approval queue — All actions require review before execution
✅ Draft replies — LLM-powered reply suggestions
✅ Unsubscribe handling — RFC 2369 compliant automatic unsubscribe

Quick Start (5 Minutes)

cd mail-ai-manager
python3 app.py  # Start Flask server at http://localhost:5051

# In another terminal:
curl -X POST http://localhost:5051/api/mail/test-connection \
  -H "Content-Type: application/json" \
  -d '{
    "mail_imap_host": "imap.gmail.com",
    "mail_imap_port": "993",
    "mail_imap_username": "your-email@gmail.com",
    "mail_imap_password": "your-app-password",
    "mail_account_name": "Gmail"
  }'

Supported Providers

Provider	IMAP Host	Port	Notes
Gmail	`imap.gmail.com`	993	Requires 16-char app-specific password
iCloud	`imap.mail.me.com`	993	Requires app-specific password
Outlook	`outlook.office365.com`	993	Use main password or app password if 2FA
ProtonMail	`127.0.0.1`	1143	Via IMAP Bridge (localhost)
Custom	Any IMAP server	Any	Any IMAP-compatible email service

API Endpoints

Endpoint	Method	Purpose
`/api/status`	GET	Check system status (Ollama, mail connection)
`/api/mail/test-connection`	POST	Test IMAP credentials
`/api/mail/status`	GET	Check Mail connection status
`/api/setup`	POST	Save IMAP config
`/api/pipeline/run`	POST	Start email classification (max 30 emails)
`/api/pipeline/status`	GET	Check pipeline progress
`/api/emails`	GET	List classified emails
`/api/actions`	GET	List pending actions (archive, trash, flag)
`/api/actions/{id}/approve`	POST	Approve action
`/api/actions/{id}/reject`	POST	Reject action

Documentation

PHASE4_QUICKSTART.md — 5-minute setup guide with curl examples
PHASE4_TEST_REPORT.md — Full integration test results + troubleshooting
MAIL_AI_SETUP.md — Provider-specific setup guides
MAIL_AI_MIGRATION_GUIDE.md — Phase 3 migration details

Architecture

Mail AI Manager (Phase 4)
├── mail_client.py
│   ├── IMAPMailClient (primary)
│   ├── AppleScriptMailClient (fallback)
│   └── HybridMailClient (orchestrator)
├── mail_action_engine.py
│   ├── fetch_unread_mail() — Get emails from INBOX
│   ├── run_pipeline() — Fetch → Classify → Route
│   ├── execute_action() — Archive, trash, flag, read
│   └── Approval queue system
├── app.py (Flask API)
├── database.py (SQLite)
└── llm_engine.py (Ollama integration)

Configuration Example

{
  "mail_imap_host": "imap.gmail.com",
  "mail_imap_port": "993",
  "mail_imap_username": "your-email@gmail.com",
  "mail_imap_password": "your-app-password",
  "mail_account_name": "Gmail",
  "ollama_model": "mistral",
  "auto_archive_spam": "true",
  "require_approval": "true",
  "auto_threshold": "0.90"
}

Performance

IMAP connect: ~500ms
Fetch 30 emails: ~2-3s
Classify 30 emails: ~10-30s (depends on LLM model)
Full pipeline (30 emails): ~15-40s

🔐 Security

✅ Passwords stored locally in SQLite (no cloud)
✅ Use app-specific passwords (not main account password)
✅ 100% local processing — no data leaves your Mac
✅ Ollama runs locally — no API calls to external services

Known Limitations

IMAP-compatible provider required
Unsubscribe requires List-Unsubscribe header in email
AppleScript fallback limited to macOS Mail.app
Local Ollama must be running separately

Troubleshooting

IMAP connection failed?

python3 << 'EOF'
import imaplib
try:
    imap = imaplib.IMAP4_SSL('imap.gmail.com', 993)
    imap.login('user@gmail.com', 'app-password')
    print("✅ Success!")
except Exception as e:
    print(f"❌ Error: {e}")
EOF

Ollama not running?

curl http://localhost:11434/api/tags  # Check status
ollama serve &                        # Start if needed

Folder Name Update

The folder has been renamed from gmail-ai-manager to mail-ai-manager to better reflect its universal IMAP provider support. All functionality remains the same.

🧊 Tool 4 — TripoSR 3D Pipeline

Generate 3D meshes from images using Apple Silicon GPU (MPS).

cd triposr-pipeline
bash setup.sh     # First time only — installs TripoSR
bash start-ui.sh
# Open: http://localhost:5050

Pipeline:

Your prompt → Stable Diffusion image → TripoSR → OBJ/GLB mesh

Uses Apple Silicon MPS (Metal Performance Shaders)
Memory usage: ~6–10GB
Output: OBJ + GLB files ready for Blender, Unity, web
Best open-source 3D model for Mac 16GB

🤖 Tool 5 — Tax AI Social ⭐ NEW

AI-powered social media content engine for tax preparation, tax resolution, and bookkeeping firms.

What it does

Automatically generates Instagram, Facebook, and TikTok posts daily at 6 AM
Uses your local Ollama LLM (mistral:7b) — no OpenAI API needed
35+ compliance rules block misleading tax claims automatically
Human review dashboard — Approve ✅, Edit ✏️, or Reject ❌ before posting
Auto-posts to Facebook/Instagram via Meta Graph API on approval
Business contact info (phone, WhatsApp, email, website) auto-appended to every post

Setup

cd tax-ai-social
bash setup.sh
cp .env.example .env
# Edit .env — add your firm info and Meta API credentials
bash start.sh
# Open: http://localhost:5055

Dashboard features

⚡ Generate Now — single post or full daily batch (5 posts)
📋 Posts tab — review drafts, approve, edit, or reject
⚙️ Business Settings tab — update firm name, phone, email, website, WhatsApp, fax, etc. without touching files
6 AM auto-generation — every morning, 5 posts ready for review
Compliance checker — flags and blocks prohibited phrases before you see them

Platforms supported

Platform	Type	Auto-posts?
Instagram	Image caption	✅ (with image URL)
Facebook	Text post	✅
TikTok	Video script	❌ Manual (record yourself)

Specialties supported

Tax Preparation
Tax Resolution (IRS debt, payment plans, OIC)
Bookkeeping

🤖 AI Models

Model	Size	Best For
`mistral:7b`	4.4 GB	General chat, posts, email, social media
`deepseek-coder:6.7b`	3.8 GB	Code generation, debugging, quant analysis

Pull additional models

ollama pull llama3.2        # Meta's Llama 3.2 (3B)
ollama pull phi4            # Microsoft Phi-4 (14B)
ollama pull qwen2.5-coder   # Qwen coding model
ollama list                 # See all installed models

🛑 Stop / Start Services

# Stop Docker services (Open WebUI + SearXNG)
docker compose down

# Start Docker services
docker compose up -d

# View logs
docker compose logs -f

# Restart a specific service
docker compose restart open-webui

📁 Project Structure

MYLLM/
├── README.md                     # This file
├── docker-compose.yml            # Open WebUI + SearXNG
├── setup.sh                      # One-command setup
├── install-tools.sh              # Install Open WebUI tools
├── shopping_search_tool.py       # Amazon/eBay search tool
├── quant_tool.py                 # Quant analysis Open WebUI tool
├── searxng/                      # SearXNG config
├── quant_api/                    # Quant AI FastAPI backend
├── mail-ai-manager/              # Mail AI email manager (IMAP)
├── triposr-pipeline/             # Image → 3D mesh pipeline
│   ├── architecture.svg          # Full stack diagram
│   └── ...
├── tax-ai-social/                # ⭐ Tax AI Social Media Engine
│   ├── app/
│   │   ├── main.py               # Flask API + routes
│   │   ├── generator.py          # Post generation engine
│   │   ├── compliance.py         # 35+ tax compliance rules
│   │   ├── poster.py             # Meta Graph API posting
│   │   ├── scheduler.py          # 6 AM daily auto-generation
│   │   └── database.py           # SQLite post tracking
│   ├── prompts/                  # 7 platform-specific prompts
│   ├── templates/index.html      # Dark-mode dashboard
│   ├── .env.example              # Config template
│   ├── requirements.txt
│   ├── setup.sh
│   └── start.sh
└── whisper-stt/                  # ⭐ Local Voice / Speech-to-Text
    ├── server.py                 # OpenAI-compatible STT API
    ├── requirements.txt
    ├── setup.sh
    ├── start.sh
    └── README.md

🔧 Troubleshooting

Ollama not connecting

ollama list          # Check if running
ollama serve         # Start if not running

Docker containers not starting

docker compose down
docker compose up -d
docker compose logs

Tax AI Social — Ollama model not found

ollama pull mistral:7b     # Download the model
ollama serve               # Make sure Ollama is running

Port conflicts

Edit docker-compose.yml or the relevant .env file to change ports.

🔒 Privacy

Everything runs 100% locally:

No data sent to OpenAI, Anthropic, or any cloud service
SearXNG proxies web searches anonymously
Gmail AI only reads emails locally via OAuth — nothing uploaded
Tax AI Social posts never leave until you click Approve
All AI inference runs on your Mac's Apple Silicon chip

🎙️ Tool 6 — Whisper STT ⭐ NEW

100% local speech-to-text for Open WebUI — talk to your AI instead of typing.

cd whisper-stt
bash start.sh
# Server starts at http://localhost:9000

First run downloads the Whisper small model (~460 MB, one time).

Connect to Open WebUI

Open http://localhost:3000
Profile icon → Settings → Audio
Speech to Text:
- Engine: OpenAI API
- Base URL: http://localhost:9000/v1
- API Key: local
Save — click the 🎤 microphone in chat to speak!

Model options

Model	Size	Speed	Best For
`tiny`	~75 MB	~0.3s	Quick testing
`base`	~145 MB	~0.5s	English-only
`small`	~460 MB	~1s	Recommended ✅
`medium`	~1.5 GB	~3s	Best accuracy

WHISPER_MODEL=medium bash start.sh   # Use a larger model

✅ 100% private — audio never leaves your Mac
✅ Works offline after first setup
✅ Multi-language support (auto-detects)
✅ OpenAI-compatible API format

📈 Quant AI — Portfolio Tracker Dashboard

Track your stock/crypto holdings with live prices, gain/loss, and allocation chart.

Open: http://localhost:8000/portfolio

Feature	Details
📊 Positions table	Ticker, shares, avg cost, current price, gain/loss, day change
🍩 Allocation donut chart	Visual portfolio breakdown (Chart.js)
👁️ Watchlist	Track tickers without holding them
⟳ Live prices	Auto-fetches via Yahoo Finance on page load
➕ Add / remove	Simple form — no file editing needed

cd quant_api
python main.py
# Open: http://localhost:8000/portfolio

📅 Gmail AI — Calendar Integration

Detect meetings in emails and add them to Google Calendar automatically.

Open: http://localhost:5051 → click 📅 Calendar in sidebar

Feature	Details
🔍 Meeting detection	Scans emails for Zoom, time mentions, meeting keywords
🤖 LLM extraction	Uses Ollama to extract date/time/location from email body
📅 One-click add	Click 📅 Add to Calendar on any detected meeting email
➕ Manual events	Create events directly from the dashboard
📋 Upcoming view	See next 7 days of Google Calendar events

Requires: Google Calendar API scope (enabled automatically when you re-connect Gmail OAuth)

⚠️ If you already connected Gmail, click 🔗 Re-connect Gmail in Settings to add Calendar scope.

🎨 Tool 7 — Stable Diffusion Image Generation

100% local text-to-image generation on your Mac. No API key. No cloud. No limits.

cd stable-diffusion
bash setup.sh     # First time only (~2-3 min)
bash start.sh
# Open: http://localhost:5050

Feature	Details
⚡ SDXL-Turbo	Default model — ~3–8s per image on M1/M2/M3
🧠 6 Models	SDXL-Turbo, SD 2.1, SD 1.5, DreamShaper, OpenJourney, Realistic Vision
📐 Any size	512×512 to 1280×1280 with 1:1, 16:9, 9:16, 4:3 presets
🎲 Seed control	Reproduce or vary any image
🗂️ History	Browse all generated images
📱 Social Post tab	Auto-generate images for Tax AI Social posts

Tax AI Social Integration

When Stable Diffusion is running (:5050), Tax AI Social will automatically generate images for Instagram and Facebook posts. Just start both services:

# Terminal 1
cd stable-diffusion && bash start.sh

# Terminal 2
cd tax-ai-social && bash start.sh

Images are auto-selected based on post content (IRS/debt → resolution style, family → family style, etc.).

API

# Generate image
curl -X POST http://localhost:5050/api/generate \
  -H "Content-Type: application/json" \
  -d '{"prompt": "professional tax accountant, modern office, 4K", "model": "sdxl-turbo", "steps": 4}'

# Use in Python
from stable_diffusion.sd_client import generate_for_post
img_url = generate_for_post("Tax deadline April 15th", specialty="tax_preparation")

🗺️ Roadmap

~~Stable Diffusion image generation~~ — ✅ Done! (stable-diffusion/ + Tax AI Social integration)
LinkedIn support for Tax AI Social
~~Quant AI portfolio tracker dashboard~~ — ✅ Done! (/portfolio)
~~Gmail AI calendar integration~~ — ✅ Done! (Calendar tab)
~~Voice interface for Open WebUI~~ — ✅ Done! (Whisper STT)

🧠 Tool 8 — AI Model Training Studio ⭐ NEW

Fine-tune and train AI models on your custom data. LoRA training, data processing, export to Ollama. Built with Streamlit.

cd ai-model-training
streamlit run app.py
# Open: http://localhost:8501

What it does

📄 Data import — Load training data from PDF, TXT, or CSV files
🎓 LoRA training — Fine-tune models without retraining from scratch
🧠 Model selection — Train on any Ollama model (mistral:7b, llama2, etc.)
📊 Training dashboard — Monitor loss, accuracy, training progress in real-time
💾 Export to Ollama — Save trained models as Ollama-compatible formats
⚙️ Hyperparameter tuning — Control learning rate, batch size, epochs, etc.
✅ Data validation — Automatic data cleaning and quality checks

Quick Start

cd ai-model-training
pip install -r requirements.txt
streamlit run app.py
# Open: http://localhost:8501

Features

Feature	Details
📚 Multi-format input	PDF, TXT, JSON, CSV training data
🎯 LoRA training	Low-rank adaptation — fast fine-tuning
🧠 Model selection	Any Ollama model supported
📈 Real-time monitoring	Loss curves, accuracy tracking
💾 Export formats	GGUF, SafeTensors, PyTorch
🚀 One-click deploy	Export directly to Ollama
📊 Data processing	Auto-tokenization, chunking, filtering

Training workflow

Your Data (PDF/TXT) → Data Processing → LoRA Training → Model Export → Ollama

Use cases

Specialized domain knowledge — Train on your company docs, legal contracts, industry papers
Custom business logic — Fine-tune to your specific use cases (tax terms, medical jargon, code style)
Faster inference — Create smaller, specialized models for specific tasks
Privacy-first — All training happens locally on your Mac

Built with ❤️ using Ollama · Open WebUI · SearXNG · TripoSR · Streamlit · Flask · Apple Silicon 🍎

Name		Name	Last commit message	Last commit date
Latest commit History 35 Commits
mail-ai-manager		mail-ai-manager
quant_api		quant_api
searxng		searxng
stable-diffusion		stable-diffusion
tax-ai-social		tax-ai-social
triposr-pipeline		triposr-pipeline
whisper-stt		whisper-stt
.gitignore		.gitignore
MAIL_AI_MIGRATION_GUIDE.md		MAIL_AI_MIGRATION_GUIDE.md
MAIL_AI_SETUP.md		MAIL_AI_SETUP.md
PHASE4_QUICKSTART.md		PHASE4_QUICKSTART.md
PHASE4_TEST_REPORT.md		PHASE4_TEST_REPORT.md
README.md		README.md
banner.svg		banner.svg
docker-compose.yml		docker-compose.yml
hub.html		hub.html
install-quant-tool.sh		install-quant-tool.sh
install-tools.sh		install-tools.sh
mail_action_engine.py		mail_action_engine.py
mail_client.py		mail_client.py
quant_tool.py		quant_tool.py
secrets.template.json		secrets.template.json
setup.sh		setup.sh
shopping_search_tool.py		shopping_search_tool.py

Folders and files

Latest commit

History

Repository files navigation

🤖 MyLLM – Personal AI Stack on Your Mac

🧰 Tools in This Stack

🏗️ Architecture

📋 Requirements

🚀 Quick Start

Step 1 — Clone the repo

Step 2 — Run setup

Step 3 — Create your account

💬 Tool 1 — Open WebUI + SearXNG

Enable web search

Install Shopping Search Tool

📈 Tool 2 — Quant AI

📧 Tool 3 — Mail AI Manager ⭐ UPDATED

What's New in Phase 4 ✨

Quick Start (5 Minutes)

Supported Providers

API Endpoints

Documentation

Architecture

Configuration Example

Performance

🔐 Security

Known Limitations

Troubleshooting

Folder Name Update

🧊 Tool 4 — TripoSR 3D Pipeline

🤖 Tool 5 — Tax AI Social ⭐ NEW

What it does

Setup

Dashboard features

Platforms supported

Specialties supported

🤖 AI Models

Pull additional models

🛑 Stop / Start Services

📁 Project Structure

🔧 Troubleshooting

Ollama not connecting

Docker containers not starting

Tax AI Social — Ollama model not found

Port conflicts

🔒 Privacy

🎙️ Tool 6 — Whisper STT ⭐ NEW

Connect to Open WebUI

Model options

📈 Quant AI — Portfolio Tracker Dashboard

📅 Gmail AI — Calendar Integration

🎨 Tool 7 — Stable Diffusion Image Generation

Tax AI Social Integration

API

🗺️ Roadmap

🧠 Tool 8 — AI Model Training Studio ⭐ NEW

What it does

Quick Start

Features

Training workflow

Use cases

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages