Real-World Use Cases

graph LR
    A[Add API Key] --> B[Start Chatting]
    B --> C[Get Results]

    style A fill:#00d4ff,color:#000
    style B fill:#141822,color:#fff
    style C fill:#00ff88,color:#000

graph LR
    A[System Prompt] --> B[SAM's Behavior]

    style A fill:#00d4ff,color:#000
    style B fill:#141822,color:#fff

graph LR
    A[You pick a personality] --> B[SAM adjusts tone and style]
    B --> C[Same power, different approach]

    style A fill:#00d4ff,color:#000
    style B fill:#141822,color:#fff
    style C fill:#141822,color:#fff

graph TD
    A[Create mini-prompts once] --> B[Enable per conversation]
    B --> C[SAM remembers your context]

    style A fill:#00d4ff,color:#000
    style B fill:#141822,color:#fff
    style C fill:#141822,color:#fff

graph LR
    A[Conversation Settings] --> B{Tools Enabled?}
    B -->|Yes| C[SAM can take action]
    B -->|No| D[Chat only]

    style A fill:#00d4ff,color:#000
    style B fill:#141822,color:#fff
    style C fill:#00ff88,color:#000
    style D fill:#141822,color:#fff

graph TD
    A[Your messages] --> D[Context Window]
    B[SAM's responses] --> D
    C[Attached documents] --> D
    D --> E[SAM sees all of this]

    style A fill:#00d4ff,color:#000
    style B fill:#00d4ff,color:#000
    style C fill:#00d4ff,color:#000
    style D fill:#141822,color:#fff
    style E fill:#141822,color:#fff

graph TD
    A[1. System Prompt<br/>Core behavior] --> B[2. Personality<br/>Communication style]
    B --> C[3. Mini-Prompts<br/>Your context]
    C --> D[4. Tools<br/>Available actions]
    D --> E[5. Conversation<br/>Current discussion]
    E --> F[SAM's Response]

    style A fill:#00d4ff,color:#000
    style B fill:#0099cc,color:#000
    style C fill:#007799,color:#fff
    style D fill:#005566,color:#fff
    style E fill:#003344,color:#fff
    style F fill:#00ff88,color:#000

flowchart LR
    A[Import PDFs] --> B[SAM analyzes]
    B --> C[Trends identified]
    C --> D[Questions prepared]

    style A fill:#00d4ff,color:#000
    style D fill:#00ff88,color:#000

You: I have lab results from 2020-2024. Help me understand 
my cholesterol trends before my cardiology appointment.

SAM: I've analyzed your lipid panels. Here's what I found:

TRENDS:
- LDL: Started at 142, now at 118 (improved)
- HDL: Stable around 52-55 (borderline low)
- Triglycerides: Spike in 2022 (285), now 168

QUESTIONS FOR YOUR CARDIOLOGIST:
1. LDL improved after starting medication in 2022. 
   Should we continue current dose?
2. HDL has been consistently in the 50s. What lifestyle 
   changes could help raise it?

DISCLAIMER: This is for informational purposes. 
Always discuss with your doctor.

flowchart LR
    A[Import contract] --> B[SAM reviews clauses]
    B --> C[Flags concerns]
    C --> D[Questions to ask]

    style A fill:#00d4ff,color:#000
    style D fill:#00ff88,color:#000

You: Review this employment contract. Explain the 
non-compete clause and flag anything unusual.

SAM: I've analyzed your employment agreement.

NON-COMPETE ANALYSIS (Section 8.2):
- Duration: 18 months post-employment
- Geographic scope: "Anywhere employer does business"
- Industry restriction: "Competing technology companies"

POTENTIAL CONCERNS:
1. California generally does NOT enforce non-competes.
   This clause may be unenforceable.
2. The geographic scope is extremely broad and 
   typically disfavored by courts.
3. 18 months is on the longer end of typical durations.

QUESTIONS TO ASK:
- Will they modify the non-compete given CA law?
- Can the scope be narrowed to direct competitors only?

DISCLAIMER: This is educational analysis, not legal advice.

flowchart LR
    A[Import statements] --> B[SAM categorizes]
    B --> C[Summary created]
    C --> D[Ready for accountant]

    style A fill:#00d4ff,color:#000
    style D fill:#00ff88,color:#000

You: I have statements from Fidelity and Schwab, plus 
crypto from Coinbase. Help me organize for my accountant.

SAM: I've analyzed your statements.

SUMMARY FOR YOUR ACCOUNTANT:

INVESTMENT INCOME:
- Dividends (qualified): $3,247
- Dividends (ordinary): $892
- Interest: $156

CAPITAL GAINS/LOSSES:
- Short-term gains: $4,521
- Long-term gains: $12,847
- Short-term losses: $2,103

CRYPTO (Needs Attention):
- 47 taxable events identified
- Net gain: ~$8,200 (needs Form 8949)
- 3 transactions may qualify as wash sales

I've created a summary document. Would you like me 
to organize by account or transaction type?

graph TD
    A[Create mini-prompts for your stack] --> B[Enable Tools and Terminal]
    B --> C[Select Tinkerer personality]
    C --> D[Ready to code]

    style A fill:#00d4ff,color:#000
    style B fill:#141822,color:#fff
    style C fill:#141822,color:#fff
    style D fill:#00ff88,color:#000

You: Add a dark mode toggle to the header component

SAM: [Reads your existing Header.tsx]
[Checks your Tailwind config]
[Creates useTheme hook following your patterns]
[Updates Header.tsx with toggle button]
[Adds dark: variants to affected components]

Done. Here's what I did:
1. Created src/hooks/useTheme.ts with localStorage persistence
2. Updated Header.tsx with a toggle button (matches your UI patterns)
3. Added dark: variants to components

Run 'npm run dev' to test.

graph TD
    A[Create Shared Topic] --> B[Create Backend conversation]
    A --> C[Create Frontend conversation]
    A --> D[Create Testing conversation]
    B --> E[All share same files and memory]
    C --> E
    D --> E

    style A fill:#00d4ff,color:#000
    style B fill:#141822,color:#fff
    style C fill:#141822,color:#fff
    style D fill:#141822,color:#fff
    style E fill:#00ff88,color:#000

You: Create a Flask API with user authentication
SAM: Created backend/app.py, backend/auth.py, backend/models.py

You: Read the backend API and create React components
SAM: [Reads backend/app.py from shared workspace]
Found endpoints: /login, /register, /products
Creating React components...

You: Write integration tests for the API
SAM: [Accesses both backend and frontend code]
Creating tests in tests/

graph LR
    A[Import papers] --> B[Search semantically]
    B --> C[Synthesize themes]
    C --> D[Write with citations]

    style A fill:#00d4ff,color:#000
    style B fill:#141822,color:#fff
    style C fill:#141822,color:#fff
    style D fill:#00ff88,color:#000

You: I have 47 papers on machine learning in healthcare. 
Help me identify main themes and gaps.

SAM: I've analyzed all 47 papers.

MAJOR THEMES:

Theme 1: Diagnostic Imaging (18 papers)
- Primary focus: radiology, pathology
- Key finding: CNN architectures dominate
- Gap: Few studies on deployment in resource-limited settings

Theme 2: Clinical Decision Support (14 papers)
- Primary focus: risk prediction, treatment selection
- Key finding: Explainability is major concern
- Gap: Longitudinal validation studies rare

Theme 3: NLP for Clinical Notes (11 papers)
- Primary focus: EHR analysis
- Key finding: Privacy-preserving methods emerging
- Gap: Limited non-English language research

RESEARCH GAPS (Potential thesis directions):
1. Intersection of explainability AND deployment
2. Cross-institutional validation studies
3. Patient-facing applications (only 3 papers!)

flowchart LR
    A[Request topic] --> B[SAM researches web]
    B --> C[Synthesizes sources]
    C --> D[Polished content]

    style A fill:#00d4ff,color:#000
    style D fill:#00ff88,color:#000

You: Research AI developments this week and create a 
newsletter summary

SAM: [Searches multiple sources]
[Scrapes relevant articles]
[Synthesizes information]

AI WEEKLY DIGEST

HEADLINE: Major breakthrough in reasoning models...

[Continues with well-structured content]

You: Create a flowchart showing user login flow

SAM: Here's the login flow:

```mermaid
flowchart TD
    A[User visits site] --> B{Has account?}
    B -->|Yes| C[Enter credentials]
    B -->|No| D[Sign up]
    C --> E{Valid?}
    E -->|Yes| F[Dashboard]
    E -->|No| G[Try again]
    D --> C

**Step 2:** SAM renders the diagram inline. You'll see a visual flowchart, not code.

**Step 3:** Ask for modifications:

### Example Requests

**Architecture diagram:**

**Database schema:**

**Project timeline:**

**Process visualization:**

### Tips for Better Diagrams

1. **Be specific**: "flowchart with 5 steps" is better than "make a diagram"
2. **Name your elements**: "boxes for Frontend, Backend, Database" gives clearer results
3. **Iterate**: Ask SAM to add, remove, or restyle elements
4. **Request different types**: Try "show this as a sequence diagram instead"

---

## LoRA Model Fine-Tuning

### Customer Support Bot

**The situation:** You want to train a model on your product documentation to answer customer questions accurately.

**Why LoRA:** Train the model to understand your product terminology, common questions, and company voice without sending data to cloud providers.

**Setup:**

1. **Prerequisites**:
   - Local MLX model installed (e.g., Qwen 2.5 7B)
   - Product documentation or FAQ documents
   - macOS with Apple Silicon

2. **Export Training Data**:
   - Go to SAM → Documents
   - Import your product documentation
   - Click **Export for Training**
   - Choose **Semantic (Paragraphs)** chunking
   - Enable PII redaction
   - Save JSONL file

3. **Configure Training**:
   - Go to Preferences → Model Training
   - Click **Train LoRA Adapter** tab
   - Choose JSONL file
   - Select base model (Qwen 2.5 7B)
   - Name: "Customer Support Bot"
   - Parameters:
     - Rank: 16 (good capacity for product info)
     - Epochs: 5
     - Batch Size: 4
     - Learning Rate: 0.0001 (default)

4. **Train**:
   - Click **Start Training**
   - Wait for completion (15-30 minutes typical)
   - Watch loss decrease (should reach ~1.5-2.0)

**Using the Adapter**:

1. Create new conversation
2. Select `lora/Customer Support Bot` in model picker
3. Ask customer questions
4. Compare responses to base model

**Example Results**:

**Benefits**:
- Accurate, company-specific responses
- No hallucination of features
- Consistent brand voice
- Complete privacy (local training and inference)

---

### Domain Expert Specialization

**The situation:** You need a specialized assistant in a technical field (medical, legal, engineering).

**Why LoRA:** Train the model on domain-specific textbooks, papers, and terminology to create an expert assistant.

**Example: Medical Terminology Assistant**

**Setup:**

1. **Collect Training Data**:
   - Medical textbooks (PDFs)
   - Clinical guidelines
   - Medical terminology references
   - Case studies (with PII removed)

2. **Export Documents**:
   - Import all documents to SAM
   - Export with **Page Aware (PDFs)** chunking
   - Enable FULL PII redaction (critical for medical data)
   - Save JSONL file

3. **Configure Training**:
   - Base Model: Llama 3.1 8B or Qwen 2.5 7B
   - Name: "Medical Terminology Assistant"
   - Parameters:
     - Rank: 32 (higher for complex medical reasoning)
     - Epochs: 10
     - Batch Size: 2 (slower, more thorough)
     - Learning Rate: 0.00005 (lower for stability)

4. **Train and Validate**:
   - Start training (may take 1-2 hours)
   - Test with medical terminology questions
   - Compare accuracy to base model

**Example Results**:

**⚠️ Important**: For educational/research purposes only. Never use for actual medical diagnosis.

---

### Writing Style Adaptation

**The situation:** You want the model to write in a specific style or voice.

**Why LoRA:** Train on examples of the desired writing style so the model consistently produces content matching that voice.

**Example: Company Brand Voice**

**Setup:**

1. **Gather Style Examples**:
   - Approved marketing copy
   - Blog posts
   - Email templates
   - Social media content
   - Company communication guidelines

2. **Create Training Examples**:
   - Format as question-answer pairs:
   ```jsonl
   {"messages": [{"role": "user", "content": "Write a product announcement for our new feature"}, {"role": "assistant", "content": "[Example in your brand voice]"}]}
   ```
   - Include 50-100 examples covering different content types

3. **Configure Training**:
   - Base Model: Qwen 2.5 3B (smaller model fine for style)
   - Name: "Brand Voice - Marketing"
   - Parameters:
     - Rank: 8 (style doesn't need huge capacity)
     - Epochs: 5
     - Batch Size: 8
     - Learning Rate: 0.0001

4. **Test Style Consistency**:
   - Generate multiple pieces of content
   - Verify tone, vocabulary, formatting match brand guidelines

**Example Results**:

**Use Cases**:
- Marketing content generation
- Email drafting
- Blog posts
- Social media content
- Internal communications

---

### Code Style Assistant

**The situation:** You want code suggestions that match your team's coding standards.

**Why LoRA:** Train on your codebase so suggestions follow your architecture patterns, naming conventions, and style guide.

**Setup:**

1. **Export Code Examples**:
   - Export well-written code files from your project
   - Include code review comments showing corrections
   - Include documentation of standards
   - Use **Fixed Size** chunking for consistent context

2. **Create Training Examples**:
   - Format as "implement feature X" → "code following your standards"
   - Include examples of common patterns
   - 100-200 examples recommended

3. **Configure Training**:
   - Base Model: Qwen 2.5 Coder 7B or Llama 3.1 8B
   - Name: "Team Code Style - Python FastAPI"
   - Parameters:
     - Rank: 16
     - Epochs: 5
     - Batch Size: 4
     - Learning Rate: 0.0001

4. **Validate Code Quality**:
   - Test with real feature requests
   - Verify code follows team standards
   - Check naming conventions, error handling, documentation

**Example Results**:

Args:
    credentials: User login credentials
    db: Database connection
    logger: Application logger

Returns:
    AuthResponse with JWT token and user info

Raises:
    HTTPException: If credentials invalid
"""
# [Follows team's error handling, logging, and structure patterns]

**Benefits**:
- Consistent code style across team
- Follows architecture patterns
- Includes proper error handling
- Matches documentation standards

---

## Business Analysis

### Customer Analysis

**The situation:** You have customer data and need insights for retention strategy.

```mermaid
flowchart LR
    A[Import data] --> B[SAM analyzes patterns]
    B --> C[At-risk identified]
    C --> D[Action recommendations]

    style A fill:#00d4ff,color:#000
    style D fill:#00ff88,color:#000

You: Analyze my customer purchase history and identify 
who I should focus retention efforts on.

SAM: Based on your customer database:

HIGH-VALUE AT-RISK CUSTOMERS (Priority Outreach):

| Customer    | Lifetime Value | Last Purchase | Risk Level |
|-------------|---------------|---------------|------------|
| Acme Corp   | $24,500       | 89 days ago   | HIGH       |
| Beta LLC    | $18,200       | 67 days ago   | MEDIUM     |
| Delta Inc   | $15,800       | 112 days ago  | CRITICAL   |

RECOMMENDED ACTIONS:
1. Delta Inc: Immediate outreach. They haven't ordered in 
   112 days but historically order every 45 days.
2. Acme Corp: Schedule check-in. Order frequency dropped 
   from monthly to quarterly in Q3.

Would you like me to draft outreach emails?

flowchart LR
    A[Request research] --> B[SAM searches web]
    B --> C[Analyzes findings]
    C --> D[Intelligence report]

    style A fill:#00d4ff,color:#000
    style D fill:#00ff88,color:#000

You: Research our top 3 competitors' recent product announcements

SAM: [Searches company websites, news, press releases]

COMPETITIVE INTELLIGENCE REPORT:

COMPETITOR A:
- Announced: New enterprise tier (March 2025)
- Pricing: $99/seat vs our $79/seat
- Key feature: Advanced analytics dashboard

COMPETITOR B:
- Announced: Mobile app launch (Q2 2025)
- Strategy: Focus on small business market
- Weakness: No API access in base tier

COMPETITOR C:
- No major announcements
- Note: Recent layoffs suggest budget constraints

STRATEGIC IMPLICATIONS:
1. Our pricing remains competitive
2. Mobile app should be prioritized
3. Opportunity to capture Competitor C's customers

graph TD
    A[What's your priority?] --> B{Privacy?}
    B -->|Yes| C[Local MLX or GGUF]
    B -->|No| D{Task type?}
    D -->|Long documents| E[Claude 3.5]
    D -->|General chat| F[GPT-4]
    D -->|Huge files| G[Gemini 1.5]
    D -->|Fast + cheap| H[GPT-4 Turbo or DeepSeek]

    style A fill:#00d4ff,color:#000
    style B fill:#141822,color:#fff
    style C fill:#00ff88,color:#000
    style D fill:#141822,color:#fff
    style E fill:#00ff88,color:#000
    style F fill:#00ff88,color:#000
    style G fill:#00ff88,color:#000
    style H fill:#00ff88,color:#000

Start: Local model for sensitive discussion
Switch: Cloud model for complex reasoning
Return: Local model for final sensitive summary