Your chatbot runs, but a reviewer asks three questions: what problem does it solve, how do you know it works, and what happens when it fails? A screenshot of a fluent answer cannot answer all three.
AI portfolio projects are strongest when they make your engineering decisions visible. Choose a bounded task, build an observable result, test it against a baseline and document its limitations. Deployment helps a reviewer try the result, but the evidence behind it matters as much as the interface.
The five briefs below are proposed learning projects. They are not claims about work LuMay has performed for customers, and completing them does not guarantee a job. Each can start small and grow as your skills develop.
Choose one project you can finish and explain
| Project | Best starting point | What it demonstrates |
|---|---|---|
| Policy-change finder | Basic application and document skills | Source comparison and evidence handling |
| Support-ticket triage assistant | Basic Python and labeled data | Classification, abstention and evaluation |
| Appointment-request sandbox | APIs and state management | Tool use, validation and safe retries |
| Forecast review desk | Data analysis and introductory ML | Time-based evaluation and error interpretation |
| AI release comparison workbench | An existing small AI app | Regression analysis and release judgment |
Pick according to the role you want and the skills you can demonstrate honestly. Udacity’s portfolio-project guide emphasizes problem framing, evaluation and communication. The project briefs here use that general learning goal while giving you concrete implementation and review tasks.
1. Build a policy-change finder with source evidence
The problem
An operations team receives a revised policy and needs to know which instructions changed. A generic summary can miss a changed condition or present a wording change as a new obligation.
Smallest useful build
Create two short, fictional policy documents with explicit version dates. Build a tool that compares them and produces a table of possible changes, the relevant passages from both versions and a review status.
Start with deterministic text comparison as the baseline. Add an LLM only where it helps group or explain the changes. The tool should distinguish added text, removed text, changed conditions and unchanged wording. It should also allow the reviewer to mark a proposed change as uncertain.
How to test it
Create an answer key before testing: a changed date, a changed eligibility condition, a moved but unchanged paragraph, and an intentionally ambiguous revision. Measure which expected changes were found and which reported changes were false alarms.
Include at least one formatting-only difference. A tool that reports every layout change as a policy change has not solved the intended problem.
What to show in the portfolio
Publish the fictional source versions, baseline output, reviewed comparison and failure examples. Explain why you used a model for one stage and ordinary comparison for another. A useful extension is to make a new document version trigger the same test set automatically.
2. Build a support-ticket triage assistant that can abstain
The problem
A small support team needs tickets routed to the appropriate queue. Incorrectly routing every unfamiliar request is worse than identifying some as needing review.
Smallest useful build
Create a synthetic ticket set with a few clearly defined categories. Record who should handle each ticket and which cases are ambiguous. Use a simple keyword or conventional classifier as a baseline, then compare an AI-assisted approach.
The application should show the proposed queue, relevant evidence from the ticket and whether human review is required. Keep routing recommendations separate from actually changing an external ticket system.
How to test it
Reserve a held-out set that you do not use to adjust prompts or rules. Include unclear tickets, multi-issue tickets and requests outside the supported categories.
Report per-category errors and the cases sent for review. Do not report a high accuracy on easy categories while hiding the small category that consistently fails. Use an appropriate classification measure from the scikit-learn evaluation documentation, and explain its relationship to routing costs.
What to show in the portfolio
Include a confusion matrix, abstention examples and a comparison with the baseline. Document how labels were created and where your synthetic data differs from real support traffic. For an extension, ask another person to review a subset and record disagreements rather than assuming your labels are unquestionable ground truth.
3. Build an appointment-request agent in a sandbox
The problem
A conversational assistant needs to collect an appointment request without inventing available times, confusing time zones or creating duplicates after a retry.
Smallest useful build
Use a mock scheduling API and fictional users. Define actions for listing availability, preparing a request and committing a confirmed booking. Keep the action that changes state behind an explicit confirmation step.
Represent the workflow as states such as collecting requirements, checking options, awaiting confirmation and completed. The model can help interpret language, but validated application logic should control which actions are permitted.
How to test it
Try a changed time zone, an unavailable slot, a request to book without required information, a cancellation before confirmation and a timeout after the scheduling service has accepted a booking.
For retries, use an operation identifier and verify whether the original request completed before creating another. Demonstrate that a repeated tool call cannot silently create multiple appointments for the same confirmed operation.
What to show in the portfolio
Show a successful conversation and a failed-service example, together with state transitions and redacted tool traces. A polished chat screen without an action record does not demonstrate reliable orchestration.
Use sandbox data and credentials. Do not connect a public portfolio demo to a real calendar or customer system without the appropriate permissions and operational controls.
4. Build a forecast review desk rather than a prediction chart alone
The problem
An operations planner needs to know when a forecast is useful and where it is unreliable. A line chart can look convincing while concealing poor performance for specific products or time periods.
Smallest useful build
Use a public dataset with a suitable license or create clearly labeled synthetic demand data. Build a simple baseline such as a previous-period or seasonal-naive forecast before adding a more complex model.
Make a review screen with forecast, observed outcome when available, error and a short explanation of the input window. Keep predictions made at a particular date distinct from later actuals.
How to test it
Use time-ordered training and evaluation periods. Fit preprocessing on the training period and avoid features that would not have been known when the prediction was made.
Report an error measure appropriate to the data and inspect slices. If values can be zero, explain why a percentage error may be difficult to interpret. Include a level shift or missing-data challenge and show how the review interface exposes it.
What to show in the portfolio
Provide the baseline comparison, data window, evaluation period and example failures. A valuable extension is to let a reviewer record an override reason, while preserving the original prediction and avoiding claims that every override is an improvement.
LuMay’s machine-learning development page provides business context for predictive-model work. Your learning project should still state its own limited data scope and avoid describing a small demo as a production forecasting service.
5. Build an AI release comparison workbench
The problem
A new prompt appears better on a few examples. The team needs to know whether it improves the application’s intended task without introducing failures elsewhere.
Smallest useful build
Take one of the projects above. Save two candidate configurations and run both against the same versioned cases. Show outputs, expected outcomes, review decisions and relevant operational measurements side by side.
Start with explicit pass conditions. For a change finder, did it detect the required condition change with the correct source passages? For a triage assistant, did it route or abstain appropriately?
How to test it
Include a deliberate regression: a configuration that improves the common case but fails an important exception. Confirm the workbench makes that trade-off visible.
If using an LLM to assist evaluation, disclose that role and keep a human-reviewed subset. Agreement between two model outputs is not proof of correctness. Separate measured API latency and cost estimates from claims about general production performance.
What to show in the portfolio
Publish a release note explaining which candidate you selected, why, and which limitations remain. It is acceptable to reject the newer candidate. A justified “do not ship” decision can demonstrate stronger engineering judgment than always presenting an improvement.
For background, LuMay’s RAG evaluation guide covers an adjacent production task; the portfolio workbench should remain a reproducible learning artifact with its own documented cases.
Give every project a reviewable evidence package
Keep a concise README with the problem, intended user, setup steps, data origin, baseline, evaluation method, limitations and demo instructions. Record model and dependency versions so another person can understand the environment.
Hugging Face’s model-card guidance offers a useful reference for documenting intended uses and limitations. Adapt that discipline to the whole application, including tools and data, rather than documenting only the model name.
For deployment, start with a restricted demo. Use synthetic inputs, keep secrets out of the repository, set reasonable usage limits and show a clear failure state. Do not publish confidential training data, private prompts or traces that reveal credentials.
LuMay’s AI training program describes guided building with RAG, agents and related application skills. Compare any training option by the work you will produce, the review you receive and the evidence you can explain afterward. Confirm current cohort details on the program page rather than relying on dates in a blog.





