Copyright (c) 2026 MindMesh Academy. All rights reserved. This content is proprietary and may not be reproduced or distributed without permission.

7.2. Quick Reference Decision Guides


🎯 Quick Reference: Responsible AI Principles

PrincipleOne-Line Tell
FairnessEquitable outcomes across groups (watch for proxy bias)
Reliability & SafetyPredictable, safe behavior — including on unexpected inputs
Privacy & SecurityProtect personal data (privacy) + defend the system (security)
InclusivenessWorks for people of all abilities/backgrounds (access)
TransparencyPeople understand how/when AI is used
AccountabilityHumans remain answerable for the system

Remember: Transparency = understandable; Accountability = answerable. Fairness = outcomes; Inclusiveness = access.


🎯 Quick Reference: Match the Workload

Scenario CueWorkload
"Is this review positive?"Sentiment analysis
"Find the company names"Entity detection
"Give me the gist"Summarization
"Transcribe this audio"Speech recognition (speech-to-text)
"Read this aloud"Speech synthesis (text-to-speech)
"What's in this photo?"Computer vision / multimodal understanding
"Where are the objects?"Object detection
"Create an image of..."Image generation
"Pull the total off this invoice"Information extraction (Content Understanding)
"Check inventory and reorder"Agentic AI

Remember: Classify by input → output. Understanding reads media; generation creates it.


🎯 Quick Reference: Model Configuration

SettingEffectFor Factual Tasks
TemperatureRandomness/creativityLow
Top-pCandidate token poolLow
Max tokensLength cap (and cost bound)Set to need
System promptPersistent rules/roleWhere durable rules go
GroundingSupply trusted factsMain defense vs. hallucination

Remember: Temperature changes style, not intelligence. Bigger model ≠ better choice.


🎯 Quick Reference: Foundry Build Patterns

GoalKey Moves
Chat clientAIProjectClient(endpoint, credential)get_openai_client()responses.create(...)
Agent clientSame + a conversation (memory) + an agent_reference (route through agent)
Extraction appChoose analyzer (prebuilt > custom) → run on content → use structured fields
Extract-then-generateContent Understanding fields → feed to model (grounding)

Remember: Deploy a model before calling it; code references the deployment name.

Alvin Varughese
Written byAlvin Varughese
Founder18 professional certifications