Activity
Mon
Wed
Fri
Sun
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
What is this?
Less
More

Owned by Frederico

Pass the PMI-CPMAI exam faster. A structured path through the 6-phase CPMAI methodology, mock exams, drills and coaching for PMs moving into AI.

PMP
16.6k
Free
Skoolers
148.5k
Free
73 contributions to PMKinetic CPMAI Academy
Data preparation is not in the data domain
Here is the trap that costs the most marks on the PMI-CPMAI exam, and it is a filing error, not a knowledge gap. Ask most candidates where data cleaning sits in the ECO and they say Domain III, Identify Data Needs. It sounds right. It is wrong. Domain III ends when you decide whether the data meets the solution needs. The cleaning, the feature engineering and the final check before training all sit in Domain IV, Manage AI Model Development and Evaluation. WHERE THE ECO PUTS IT Domain IV, Task 4 is "Manage data transformation to conduct data preparation". Its five enablers: oversee data cleaning and preprocessing workflows. Coordinate feature engineering and selection activities. Manage data normalization and standardization processes. Supervise data augmentation and synthetic data generation. Ensure data transformation reproducibility and documentation. Domain IV, Task 5 is "Verify data quality for go/no-go decision to conduct data preparation". Its first enabler is to conduct final data quality assessments before model training. This is the second of the three go/no-go points in the outline. THE THREE GATES, IN ORDER Domain III, Task 8: does the raw data meet the solution needs? If no, go back to the business problem or find more data. Domain IV, Task 5: after preparation, is the transformed data fit to train on? If no, go back to preparation. Domain IV, Task 6: is the model ready for operationalization? If no, go back to development or evaluation. Notice that two of the three gates score under Domain IV, and only one of them is about the model. WHAT IT LOOKS LIKE ON A REAL PROJECT An insurer prepares claims data for a fraud model. The raw data passed the Domain III check. During preparation the team drops every claim with a missing adjuster code, which removes one entire regional office that uses a different system. The inputs were valid. The transformation changed the population. That is exactly what the Task 5 check exists to catch, and it is why validating inputs before transformation is not the same as validating the result after it.
1
0
Question of the Day: The Map With Three Districts Missing
A city council wants a model that ranks citizen reports of road damage so repair crews go to the worst spots first. Reports come through a smartphone app launched two years ago. During data understanding the team finds that 70% of the reports come from three of the city's twelve districts, the ones with the highest incomes and the most app users. The other nine districts, where the roads are known to be older, are almost absent from the data. The sponsor says: "The model ranks potholes, not people. Fairness does not apply here. Move on to data preparation." What should the project professional do first? A. Proceed to data preparation and record the district imbalance as a known limitation in the final project report. B. Record the imbalance as a data gap in the readiness assessment and agree how the nine districts will be covered before preparation starts. C. Ask the data scientists to apply a bias mitigation technique during training so the model weights all districts equally. D. Test the trained model's rankings across the twelve districts before go-live and correct any disparity found at that point. Reply with the letter and one line on why. One letter is fine too. The answer and the ECO tasks behind it are in the first comment, so reply before you open it.
0 likes • 11d
ANSWER: B. WHY. You are still in data understanding. The ECO gives this moment its own task: Domain III, Task 7, Oversee data evaluation, with the enabler "analyze data distributions and identify potential biases or gaps". The finding then feeds Domain III, Task 8, Determine if data meets solution needs, where the go/no-go on data readiness is made. A dataset that barely contains nine of twelve districts does not meet the needs of a model that must rank damage across all twelve. The gap is fixed at the source: pull in other report channels, use inspection records from the quiet districts, or sample them directly. That decision is made before anyone cleans or transforms anything. WHY THE SPONSOR IS WRONG. The model does not rank people, but it decides who gets a road repaired. Domain I, Task 3, Conduct bias checks, starts with "analyze training data for demographic and representation imbalances". The people affected do not have to be in the model for the outcome to be unequal. WHY NOT A. Accepting the gap now and writing it down later turns a known data problem into a production problem. WHY NOT C AND D. Both act after the model exists. Mitigation during training and fairness testing before go-live are real Domain I work, but they treat a symptom of data the model never saw. You cannot reweight districts that are not in the data. THE TRAP. Bias appears in two domains. If the output of the work is a go or no-go on the data, think Domain III. If the output is a fairness finding about the model, think Domain I.
Query
where can i find the Cheat sheet, 10-week study plan and resources
0 likes • 11d
Siddharth, all three are in the Community tab under the Resources & Templates category. Click that category in the filter row at the top of the feed and you will see them: CPMAI Exam Cheat Sheet (ECO Sept 2025), the whole exam on three pages. CPMAI 10-Week Study Plan (5 to 7h/week), one week per phase with an exit checkpoint each week. CPMAI Verified Resource List, only the official PMI links. While you are there, open The CPMAI Domains x Phases Map as well. It puts all 37 ECO tasks on the six phases and it is the single page most members say helped them most. If any file does not open for you, tell me here and I will send it another way. Which week of the plan are you starting on?
After go-live: who watches the model on day 31?
Many students tell me Domain V feels abstract. The tasks read like a list: oversee model governance, oversee AI solution metrics, prepare the final report. On the exam they turn into very concrete situations, usually set weeks or months after deployment. Here is a pattern I see a lot. A retailer deploys a demand forecasting model. Accuracy at go-live is fine. Three months later, store managers stop trusting the forecasts, because a competitor opened nearby and buying patterns changed. Nobody noticed for weeks, because the dashboard only showed system uptime. Three questions usually decide the answer in these scenarios. 1. Who owns the model now? After the transition, someone in operations must own performance, drift checks and the retraining decision. If the stem tells you nobody owns it, fixing ownership is often the first step. 2. What is being measured? Uptime and response time are not enough. You need the model metric that was agreed (forecast error, recall, and so on) and the business KPI the model was meant to move, each with a threshold and an alert. 3. What happens when a threshold is breached? A good governance plan says in advance who gets the alert, who decides between retraining, rollback or a fallback process, and how the change is versioned and approved. A common trap is the option that says "retrain the model immediately". Retraining is a decision, not a reflex. First confirm what changed: the input data, the relationship between inputs and outcome, or a broken pipeline. Then follow the change control the team agreed. Exam link: Domain V, Task 3 (Oversee model governance) and Task 4 (Oversee AI solution metrics). Question for you: in your own projects, who owns a model after it goes live? Is it written down anywhere?
1
0
Question of the Day: The Assistant Everyone Likes
A bank built an internal assistant that answers staff questions about lending policy. It retrieves passages from the policy manual and writes an answer with a link to the section it used. After a four-week pilot, 40 loan officers rate it 4.6 out of 5 and say it saves them about 20 minutes a day. Before go-live, a compliance analyst checks 50 answers by hand. In 6 of them, the linked section does not support what the answer says. Two of those 6 would have led to a wrong approval decision. The sponsor wants to go live next week, because users are happy and the pilot hit its time-saving target. What should the project professional do next? A. Go live on schedule with a clear notice telling staff to check every answer against the linked section. B. Extend the pilot to 200 loan officers so the team has a larger sample before the go-live decision. C. Hold the go-live and agree with compliance a test set and a minimum score for source-backed answers. D. Move the project to fine-tuning the language model on the full manual to remove the unsupported answers. Reply with the letter and one line on why. One letter is fine too. The answer and the ECO task behind it are in the first comment, so reply before you open it.
0 likes • 19d
Answer: C. The pilot measured satisfaction and time saved. Neither tells you whether the answers are correct. Compliance found 12% of sampled answers not backed by their source, and some would have caused wrong approvals, so the model is not ready for the go/no-go decision. The project professional makes the readiness criteria explicit: a representative question set, a score for answers backed by their sources, and a threshold the business owner (here, compliance) agrees to. Then the team decides. Why not the others: A puts the control on users who already trust the tool. A notice does not fix a 12% error rate on regulated decisions. B grows the wrong measurement. More satisfied users will not show whether the answers match the policy. D jumps to a technical fix before the problem is measured. In a retrieval-based assistant, unsupported answers often come from retrieval or prompting, and fine-tuning does not guarantee the answer matches the cited section. ECO: Domain IV, Task 6 (Verify model ready for operationalization go/no-go decision), evaluate model performance against established success criteria.
1-10 of 73
Frederico Peixoto
3
36 points to level up
@frederico-peixoto-6763
Pass PMP, PMI-ACP, CAPM and CPMAI. Structured study plans, mock exams, situational drills, coaching and accountability from diagnostic to exam day.

Active 2d ago
Joined Jul 22, 2026