ai-project-lifecycle

Source Metadata

Original cangjie-skill frontmatter from the distillation run:

id: f18
title: "AI项目生命周期 — 从立项到落地的完整流程"
book: "AI for Everyone"
author: "Andrew Ng"
source: p08, p09, p10
tags: [ai-project-lifecycle, project-management, execution, workflow]
related_skills:
  - slug: ml-workflow, relation: composes-with
  - slug: triple-due-diligence, relation: depends-on

R — Original Quote (Reading)

"The lifecycle of an AI project includes: clarifying business goals, collecting data, training models, deploying to production, continuous monitoring and iteration. Each stage requires different skills and team collaboration." (p08-p10 synthesis)

I — Interpretation

An AI project is not a purely technical project, butbusiness + technologyfusion. The complete lifecycle includes six stages:

1. Clarify business goals: What business problem should AI solve? What are the success metrics?
2. Data collection and preparation: Acquire, clean, and label data — this is usually the most time-consuming stage
3. Model training and validation: Choose a suitable algorithm, train the model, and evaluate it with a validation set
4. Deployment to production: Integrate the model into the production environment and connect it with existing systems
5. Continuous monitoring: Track model performance and detect data drift and performance degradation
6. Iterative optimization: Continuously improve based on feedback and data

Key insights:

• Each stage requiresdifferent skills(business analysis, data engineering, ML engineering, DevOps)
• Data preparation often accounts for 60-80% of project time
• Deployment is not the endpoint, but the starting point for continuous iteration

A1 — Past Application

Case 1: Intelligent customer service system An e-commerce company used AI to replace part of its human customer service. Project timeline: 2 weeks to clarify business goals (handle 80% of common questions) → 4 weeks to collect and label historical conversation data → 3 weeks to train intent recognition and answer generation models → 2 weeks to deploy to the customer service system → continuous monitoring of answer accuracy and iteration.

Case 2: Manufacturing defect detection A factory used computer vision to detect product defects. The key challenge was data collection — a large number of defect samples were needed, but defects are rare in actual production. The team solved this with data augmentation and synthetic data, and ultimately deployed real-time detection on the production line.

Case 3: Financial risk control model A bank built a loan default prediction model. The project focused on feature engineering and compliance review — the model not only had to be accurate, but also interpretable and compliant with regulatory requirements. After deployment, it continuously monitored the model's performance across different demographic groups to ensure fairness.

A2 — Future Trigger

Trigger scenarios:

• Starting a new AI/ML project
• Evaluating the progress and risks of an AI project
• Planning the workflow of an AI team
• Reporting an AI project plan to management

Language signals:

• "AI project process", "machine learning project", "model deployment", "data labeling", "model monitoring", "project milestones"

Key distinctions:

• Is thisfrom 0 to 1a new project oriterative optimizationof an existing system?
• What is the project'smain bottleneck? (Data? Algorithm? Deployment? Business understanding?)

E — Execution

1. Define business objectives: Clarify what problem AI should solve, success metrics (accuracy, ROI, user experience, etc.), and project scope
2. Assess data availability: Is existing data sufficient? What new data needs to be collected? How is the data quality?
3. Develop a project plan: Estimate time for each phase (data preparation is usually the longest), and set milestones
4. Build the team: Ensure talent across business analysis, data engineering, ML engineering, and DevOps
5. Iterative development: Start with an MVP (minimum viable product) to quickly validate feasibility, then gradually refine
6. Deployment and monitoring: Build a model monitoring dashboard and set alerts for performance degradation
7. Continuous iteration: Continuously optimize the model based on production feedback and data

Completion criteria: Complete the project plan, specifying deliverables, timelines, owners, and success metrics for each phase.

B — Boundary

Not applicable scenarios:

• Pure research exploration (no clear business objective)
• Routine optimization of an existing mature process (no need for a full lifecycle)
• Scenarios where data privacy or compliance does not allow the use of AI

Failure modes:

• Skipping business objective definition and jumping straight to technical implementation ("holding a hammer looking for nails")
• Underestimating data preparation time, causing project delays
• No monitoring after deployment, so model performance degradation is not discovered in time
• Over-pursuing model complexity while neglecting business value

Author blind spots:

• In real projects, the phasesare not linear—they often require returning to the previous phase
• Organizational collaboration and communication are often harder than technology—cross-department collaboration and managing management expectations
• Many AI projects fail not because of technology, but because ofunclear business objectivesorinsufficient organizational readiness

Related skills

• ml-workflow — composes-with: The "model training and validation" phase in the AI project lifecycle follows the ML workflow (collection → training → deployment → iteration). Together, the two cover the complete process from project initiation to technical execution.
• triple-due-diligence — depends-on: Before starting an AI project, you need to first verify through triple due diligence that the project is worth investing in. Lifecycle management is built on the basis that the project selection is correct.