1. Model Deployment Options (What are your choices?)
You’re choosing where and how the model runs.
☁️ Cloud Deployment
Use models hosted in platforms like Microsoft Azure (e.g. Azure AI services)
When to use:
You want scalability
Minimal infrastructure management
Quick to implement
Examples:
Azure OpenAI (LLMs)
Azure AI Vision
Azure AI Language
✅ Exam takeaway:
Default / most common answer unless constraints say otherwise
🖥️ Edge Deployment
Model runs on local devices (IoT, mobile, on-prem hardware)
When to use:
Low latency required
Limited/no internet access
Privacy concerns (data stays local)
Example scenario:
A factory needs real-time defect detection without internet
✅ Use: Edge deployment
🔒 On-Premises Deployment
Model runs in your own data centre
When to use:
Strict regulatory or compliance requirements
Sensitive data cannot leave organisation
🌐 Hybrid Deployment
Combination of cloud + edge/on-prem
When to use:
Balance performance, cost, and compliance
⚙️ 2. Model Configuration Parameters (How the model behaves)
These are settings that influence output quality, creativity, and control—especially important for generative AI models.
🎛️ Temperature
Controls randomness / creativity
Low (e.g. 0.2) → deterministic, predictable
High (e.g. 0.8) → creative, varied
Example:
Financial report generation
✅ Use: Low temperature
🎯 Max Tokens
Limits the length of the response
Prevents overly long outputs
Controls cost and performance
🧠 Top-P (Nucleus Sampling)
Controls probability distribution of word choices
Low → safer, focused output
High → more diverse output
🔁 Frequency Penalty
Reduces repetition of words/phrases
🚫 Presence Penalty
Encourages introducing new topics
🛑 Stop Sequences
Defines where the model should stop generating text
⚖️ 3. Matching Deployment + Parameters (Exam Thinking)
Example 1
A chatbot must give consistent, accurate answers
Deployment → Cloud (Azure OpenAI)
Temperature → Low
Max tokens → Controlled
Example 2
Real-time image processing in a remote location
Deployment → Edge
Parameters → Not the focus (vision model config is simpler)
Example 3
Sensitive healthcare data must not leave organisation
Deployment → On-premises
🧩 4. Key Exam Pattern
AI-901 questions usually test:
👉 Deployment choice based on constraints:
Latency → Edge
Compliance → On-prem
Scalability → Cloud
👉 Parameter choice based on output need:
Accuracy → Low temperature
Creativity → High temperature
Cost control → Limit tokens
🧠 5. Simple Memory Rule
“Where does it run + how does it behave?”
Where → Deployment (Cloud / Edge / On-prem / Hybrid)
How → Parameters (Temperature, Tokens, Top-P, etc.)
✅ Summary
To identify appropriate deployment and configuration:
Understand constraints (latency, compliance, scale)
Choose deployment model
Adjust parameters based on output requirements
Match scenario → capability → configuration