Author: Manish Tiwari

1. AI platform research

First, you research different AI platforms and compare what each one can do.

Examples include:

• ChatGPT — general-purpose AI for reasoning, writing, coding, research, etc.
• Claude — strong for long documents, reasoning, coding and structured work.
• Gemini — Google’s AI ecosystem, with strong multimodal and Google-service integration.
• Microsoft Copilot — AI integrated into Microsoft’s productivity and development ecosystem.
• Perplexity — focused strongly on AI-assisted search and research.
• Specialized AI platforms — platforms designed for coding, marketing, data analysis, design, customer support, and other specific jobs.

The research shouldn’t just ask “Which AI is best?” Instead, you create criteria such as:

Criteria	What you investigate
Accuracy	Does it give correct answers?
Reasoning	Can it solve complex problems?
Speed	How quickly does it respond?
Cost	Free, subscription, API costs
Context	How much information can it handle?
Multimodal	Text, image, audio, video?
Privacy	How is user/company data handled?
Reliability	Does performance remain consistent?
Explainability	Can you understand why it produced an answer?
Integration	Can it connect to other systems?

This produces a platform comparison report.

---

2. What is PACE?

There isn’t one universal AI framework called PACE. One current PACE framework defines it as:

P — Plan
A — Architect
C — Code
E — Evaluate

It uses specialized AI agents to divide software-development work into stages rather than asking one AI to do everything. PPACE Framework

Think of it like this:

Plan → Architect → Build → Test/Evaluate → Improve

For example, suppose you want AI to develop a traffic-management application.

P — Plan

Define:

• What problem are we solving?
• Who will use the system?
• What data is required?
• What are the success criteria?

A — Architect

Design:

• AI model
• database
• sensors
• APIs
• user interface
• security
• data flow

C — Code

Build the actual system.

E — Evaluate

Test:

• Accuracy
• Safety
• Bias
• Performance
• Reliability
• Failure cases

This is useful because AI development becomes a controlled process instead of “ask AI to build everything.”

There is also another current product called Pace that provides AI skills/workflows for different organizational roles such as engineering, sales, finance, legal, data, marketing and design, so you need to identify which PACE your project means. 

---

3. What is CRAFT?

Again, CRAFT is not a single universal AI acronym.

One current CRAFT framework is designed around making AI work repeatable, auditable and disciplined rather than relying on remembered prompts. It uses things such as versioned workflows, confidence gates, backups and multi-persona validation. CCRAFT Framework

The basic idea is very important for AI research:

Don’t just ask an AI a question once. Create a repeatable process that another person can inspect and reproduce.

For example:

Input → AI processing → Validation → Human review → Output → Audit record

That makes AI much more suitable for business, government, research and safety-critical applications.

---

4. What is a traffic-light audit?

This is probably the easiest concept to understand.

A traffic-light audit evaluates something using:

🟢 Green = acceptable / low risk
🟡 Yellow = needs attention / medium risk
🔴 Red = unacceptable / high risk

For AI, you could audit several categories:

Area	Result	Meaning
Accuracy	🟢	Good performance
Bias	🟡	Some potential concerns
Privacy	🟢	Appropriate controls
Explainability	🟡	More documentation needed
Security	🔴	Serious vulnerability
Human oversight	🟢	Human approval exists
Reliability	🟡	More testing required

There are published AI-audit frameworks that use traffic-light risk assessment alongside multiple technical KPIs, including model performance, discrimination/fairness, transparency and explainability. 

So a traffic-light audit is essentially a visual risk-management system.

---

5. Traffic-light audit for an AI system

Imagine you’re evaluating an AI system used to control traffic lights.

You might investigate:

Technical

• Is the AI accurately detecting vehicles?
• Can it recognize pedestrians?
• Does it work at night?
• Does it work during rain?
• What happens if a camera fails?

Safety

• Can the AI accidentally keep a direction green?
• Is there a safe fallback?
• Can humans override it?
• What happens when the AI makes an uncertain prediction?

Fairness

• Does it work equally well in different locations?
• Does it disadvantage pedestrians?
• Does it behave differently in high-traffic vs low-traffic areas?

Security

• Can someone manipulate the sensors?
• Can someone interfere with the AI?
• Is the system isolated from unauthorized access?

Governance

• Who is responsible if the AI makes a bad decision?
• Are decisions logged?
• Can investigators reconstruct what happened?

Then you could produce:

Overall AI Risk: 🟡 YELLOW

because perhaps the system performs well but still needs stronger security and human override mechanisms.

---

6. How all of these fit together

You can think of your research project as:

AI Platforms
↓
Compare different AI technologies
↓
PACE
↓
Plan → Architect → Build → Evaluate
↓
CRAFT
↓
Make the AI workflow repeatable, documented and auditable
↓
Traffic-Light Audit
↓
Green / Yellow / Red risk assessment
↓
Final Recommendation

This is a strong structure for an AI research / AI governance / AI evaluation project.

Example final research question

“How can different AI platforms be systematically evaluated using structured frameworks such as PACE, CRAFT, and traffic-light risk auditing?”

Then your research could compare ChatGPT, Claude, Gemini, Copilot, Perplexity and specialized AI systems, evaluate them against common criteria, and finally assign risk ratings.

One particularly important modern research principle is human-in-the-loop evaluation: recent Microsoft Research work found that developers want AI to take on substantial assembly work while maintaining explicit authority boundaries, provenance, uncertainty signals and human control.