Portfolio Review Queue with Generative AI: A 2027 Project Guide

Use a familiar campus workflow to practise Generative AI: portfolio review queue. The user is a peer-review organizer, and the first question is: What is the average time spent on completed portfolio reviews? The deliverable is an evidence-grounded assistant prototype with a visible rule and repeatable evidence behind its output.

This is a 2027 study-edition project brief. All example records and operational situations are fictional. Start with basic Generative AI knowledge, the ability to inspect structured data and a local practice environment. The complete first version should answer the stated question before adding accounts, real integrations or additional screens.

Define the project rule

Average completed review duration is an operations metric, not a judgement of the learner. A withdrawn review must not count as a completed assessment. For this exercise, one row represents one review. The stable key is record_id; project_category is the grouping or filtering label. The numeric review_minutes field uses minutes. Accepted states are queued, completed, withdrawn.

The summary selects status completed and applies mean to the selected numeric values. Keep this definition visible in the interface, response or report so another learner can reproduce it. Use positive finite values in the starter fixture and define any future zero or missing-value policy explicitly.

Choose the Generative AI approach

Use the application data as evidence and keep calculations in deterministic code. Ask a language model to explain a supplied result rather than inventing the result from an instruction. Treat record text as untrusted content, make unsupported questions visible, and evaluate against a small human-written answer set before describing the assistant as useful.

The three design decisions to explain in your project notes are:

• Deterministic calculations outside the model. Apply this choice to review records and explain its effect on review_minutes.
• Evidence references in answers. Apply this choice to review records and explain its effect on review_minutes.
• Explicit abstention when data is absent. Apply this choice to review records and explain its effect on review_minutes.

Work through a small sample

The three records below are deliberately small enough to inspect by hand. They are a compact specification for the data shape and expected result, and can later be expanded into a larger synthetic fixture.

[
  {
    "record_id": "portfolio-reviews-01",
    "project_category": "Web Application",
    "review_minutes": 25,
    "status": "completed",
    "recorded_on": "2027-01-04"
  },
  {
    "record_id": "portfolio-reviews-02",
    "project_category": "Data Report",
    "review_minutes": 40,
    "status": "queued",
    "recorded_on": "2027-01-05"
  },
  {
    "record_id": "portfolio-reviews-03",
    "project_category": "Mobile Prototype",
    "review_minutes": 15,
    "status": "completed",
    "recorded_on": "2027-01-06"
  }
]

Only the first and third records match completed. Their review_minutes values are 25 and 15; the full-fixture result is 20 minutes, with 2 matching records. Selecting project_category = Web Application leaves the first record, so the result becomes 25 minutes, with 1 match. This provides two independent checks for the filter and aggregation.

The full-fixture output contract is:

{
  "matching_count": 2,
  "operation": "mean",
  "value": 20,
  "unit": "minutes"
}

Check the result by project_category

The overall figure answers the all-records version of the question. Grouping by project_category produces the following separate expectations, using the same completed rule. The middle group has a record in the source fixture but no qualifying observation.

Use this breakdown when answering the group-level question. Compare like-for-like units and keep the matching count visible, especially for averages. A different filter changes which records enter the calculation; it does not change the meaning of the metric.

The code blocks describe sample data and an expected result. Implement the surrounding feature using the installed version of your chosen tools, then save the actual run evidence separately from these expectations.

Build the feature in three passes

1. Calculate before asking for prose

Use ordinary code to select completed review records and calculate the mean of review_minutes. Pass the verified result, units and matching record identifiers to the assistant. The model should explain supplied evidence rather than invent a numerical answer.

2. Define the answer and abstention format

Ask for a brief answer to the project_category question with references to the supplied records. Include an explicit outcome for absent or inconsistent data. Treat labels and other record text as data, including any instruction-like text inside a fictional record.

3. Evaluate a small answer set

Write expected answers for the full fixture, one project_category filter, no matching records and an unsupported question. Compare numbers deterministically and review evidence references separately. Record observed failures before adding more prompts or calling the prototype reliable.

Find the mistake before polishing the result

A useful defect to introduce deliberately is: Withdrawn reviews are assigned a fabricated zero duration. First reproduce it against the fixture, then identify the smallest rule or state transition that explains the mismatch. Fix that boundary and rerun the full-fixture and selected-label examples. This creates a stronger explanation than changing several unrelated parts until the screen looks right.

The scenario-specific acceptance check is: Leave withdrawn duration outside the completed-review average. Preserve before-and-after evidence. Add a missing project_category, a negative review_minutes, an unrecognised status and a duplicate record_id as separate invalid cases. Define the rejection or duplicate-handling policy before testing, and verify that invalid input does not silently change valid existing records.

Check the Generative AI concerns as well:

• The answer must not invent a missing record.
• Text inside a record must not override the task.
• A numerical answer must match the verified calculation.

For an empty filter match, the expected mean value is undefined (no matching observations). Keep matching_count at zero. This distinction prevents an empty dataset from being interpreted as a measured observation.

Explain the finished work in a portfolio

Show the peer-review organizer question, the record definition and the expected calculation first. Then demonstrate a project_category filter and the acceptance check. Include the fixture, setup instructions, chosen tool versions, actual test output and one remaining limitation in the repository. If you show a screenshot, pair it with the input that produced it. Another learner should be able to repeat the same result from a clean checkout.

For a focused extension, add a second project_category selection and compare the two groups using the same rule. Specify how missing groups appear and keep the original fixture as a regression case. Expand scope only after the current calculation, state handling and recovery path remain understandable.

Questions learners often ask

Which review records belong in the result?

Select records whose status is exactly completed and that satisfy the chosen project_category filter. Other states remain in the source data and can have separate summaries. Follow the domain rule above when a record changes state.

What should happen when no records match?

Treat the empty result as undefined and show a clear no-observations message. Keep the applied filter visible so the user can change it. Avoid using a loading indicator for a completed request that simply has no matching data.

How can I prove the review_minutes result is correct?

Reconcile the first and third sample records, then repeat the calculation for Web Application alone. Save both expected and actual outputs, and rerun the scenario-specific acceptance check after making a change.

Continue learning with Groot Academy

Use this exercise as a concrete discussion point when choosing a learning path. Browse the current Groot Academy course directory and ask about the current syllabus, learning mode and project practice. Course availability and batch details should be confirmed directly with the academy.

For implementation details, consult the official Generative AI documentation matching the tools you install. This project defines an original practice brief; it does not depend on a forecast about 2027 software releases.

Continue with Peer Feedback Tracker using Generative AI, or compare Portfolio Review Queue using Python. These companion practice guides are part of the 2027 learning collection.