Interview: A High Schooler’s AI Science Fair Project—Timeline, Mistakes, Judges’ Feedback

Meet the student: “I didn’t start with AI—I started with a question”

Parents often ask us for ai science fair project ideas high school students can actually finish—and that judges will take seriously. So I sat down (virtually) with Jordan Lee, a 16-year-old junior who built a machine learning science fair project that earned a top award at his regional fair.

Jordan’s project: “Can a model predict early signs of plant leaf disease from photos?”

He trained a simple image classifier to distinguish healthy leaves from early-stage disease symptoms. The twist: he wasn’t trying to “build the smartest AI.” He was trying to answer a measurable science question: does adding a consistent lighting setup improve model accuracy more than adding more training images?

Why this interview matters if you’re a parent:

• It shows exactly how to do AI project for science fair without needing a PhD.
• It reveals the real mistakes (data problems, time problems, presentation problems).
• It shares the kind of feedback judges give—so your teen can build smarter from day one.

The timeline: what Jordan did each week (and what he’d change)

Jordan’s biggest surprise wasn’t coding. It was planning.

“I thought I needed months of hardcore programming. Actually, I needed a week of thinking clearly about the experiment.”

Below is Jordan’s real build timeline. If your teen is starting from scratch, this is a strong template to copy.

Week	Goal	What Jordan did	Deliverable to finish that week	Parent support that helps (a lot)
1	Define the question	Chose plants, defined “early disease,” set measurable variables	1-page project plan + hypothesis	Help your teen narrow scope: one question, one metric
2	Collect baseline data	Took 300 photos in consistent lighting; labeled them	Labeled dataset v1 (folders or CSV)	Provide a quiet space + time blocks for photo sessions
3	Build first model (ugly version)	Used a beginner-friendly notebook template; trained a simple classifier	Model v1 + baseline accuracy	Encourage “done, not perfect” so they can iterate
4	Run the experiment	Changed one variable at a time (lighting vs. more images)	Results table + graphs	Help them keep a lab notebook and date everything
5	Improve reliability	Fixed labeling errors; balanced categories; re-ran training	Model v2 + confusion matrix	Ask: “What could make results unfair?”
6	Write + present	Drafted abstract, methods, results, limitations; built display	Final report + poster + demo	Do a mock judging interview at home

Jordan’s “do-over” if he had one extra week:

• He’d collect better negative examples (healthy leaves with weird shadows) earlier.
• He’d write his Methods section while building instead of at the end.

The project build: data, model, and the “science” part judges care about

A lot of students hear “AI science fair” and think the goal is: make AI. Judges usually think the goal is: do science using AI. That means variables, controls, and clear evaluation.

Step 1: Data collection (the least glamorous, most important part)

Jordan didn’t download a giant internet dataset. He made his own, because it matched his experiment.

What he photographed:

• One plant species (to reduce variables)
• Two classes: healthy vs. early disease
• Same phone camera
• Same distance and angle

He also tracked metadata in a simple spreadsheet:

• Date/time
• Lighting condition (natural window light vs. ring light)
• Location (kitchen vs. garage)

“My first dataset was messy. Some ‘healthy’ photos had yellow spots that were actually early disease. That destroyed my accuracy and I didn’t know why.”

Actionable tip for parents: if you want to help without “doing the project,” support the environment.

• Provide consistent lighting (even a cheap ring light)
• Encourage short sessions (15–20 minutes) so labeling doesn’t get sloppy
• Remind them to back up photos (Google Drive works fine)

Step 2: The model (simple beats fancy)

Jordan used a beginner-friendly image model approach (transfer learning). In plain terms: he used a model that already learned basic visual patterns and trained it to recognize his two categories.

What he measured:

• Accuracy (overall)
• Precision/recall (to show tradeoffs)
• Confusion matrix (to show where it failed)

He also tested “real life” performance:

• Photos taken in different rooms
• Slightly different angles
• Leaves with dust or water spots

“The model looked great in my training notebook, then totally failed when I took a photo at night. That’s when I realized my experiment needed a ‘real-world test’ section.”

Step 3: Making it a real experiment (not just a demo)

Jordan’s independent variable: data collection setup

• Condition A: more photos, inconsistent lighting
• Condition B: fewer photos, consistent lighting

Dependent variable: model performance (accuracy, precision/recall)

Control elements:

• Same plant type
• Same camera
• Same labeling rules

He learned something judges love: a negative result is still a result.

“I assumed more photos would always win. But consistent lighting improved accuracy more than doubling the dataset—at least for my setup.”

If your teen is searching for ai science fair project ideas high school level, this is a strong pattern:

• Pick a real-world classification problem (images, text, or sensor data)
• Choose one variable to test (data quality, feature choice, or preprocessing)
• Evaluate with clear metrics and limitations

Mistakes Jordan made (so your teen doesn’t repeat them)

This is the part most families need. Because the biggest time-wasters in a machine learning science fair project aren’t “hard coding”—they’re preventable missteps.

Mistake 1: Labeling rules were unclear

Jordan’s first labels were based on “vibes.” Then he created a simple rule:

• Healthy = no visible spots or yellowing
• Early disease = at least one visible spot or patch across two photos

Fix your teen can copy:

• Write labeling rules like a mini contract
• Label a small batch (30–50) first
• Re-check labels after the first model run

Mistake 2: He trusted accuracy too early

“My accuracy was 92% and I was proud… until I realized the model was predicting ‘healthy’ for almost everything.”

What happened: the dataset was imbalanced.

Fix:

• Balance classes (similar number of healthy vs. disease images)
• Report multiple metrics:
  • Precision
  • Recall
  • Confusion matrix

Mistake 3: He didn’t document versions

He changed the dataset, changed labels, changed the model—then forgot which combo produced which results.

Fix:

• Create a simple versioning habit:
  • dataset_v1, dataset_v2
  • model_v1, model_v2
  • results_v1.csv

Mistake 4: He overbuilt the “cool factor”

He spent a weekend trying to build a fancy app interface.

Judge reality:

• Judges usually reward clear methodology over polished UI.

Better use of time:

• One clear demo image set
• One clean results chart
• One honest limitations section

Judges’ feedback: what they praised, what they challenged

This is the gold. Jordan wrote down judge comments right after presenting.

What judges liked

• A clear hypothesis (not just “I made an AI”)
• Controlled comparison (lighting consistency vs. dataset size)
• Failure analysis (he showed examples where the model was wrong)
• Ethics and limits (he explained it’s not medical diagnosis; it’s a learning tool)

What judges pushed on (and how Jordan answered)

Common judge questions and strong answers your teen can practice:

• “How do you know your labels are correct?”

  • Jordan: “I used a labeling rule and re-checked ambiguous images. I also tracked uncertain labels separately.”

• “Could your model be overfitting?”

  • Jordan: “I tested on photos taken on different days and in different rooms. Performance dropped in low light, so I documented that limitation.”

• “What would you do next if you had more time?”

  • Jordan: “Add more plant species, test generalization, and compare a second model type.”

• “Why is this science and not just engineering?”

  • Jordan: “Because I tested a variable (lighting control) and measured its effect on outcomes with repeatable steps.”

Parent tip: do a 10-minute mock interview at home. Ask the four questions above. Help your teen answer with:

• One sentence summary
• One data point
• One limitation

Next Steps: how your teen can start an AI science fair project this week

If you’re searching how to do ai project for science fair, here’s a practical, low-stress launch plan. The goal is momentum without overwhelm.

1) Pick a question that fits in one sentence

Good templates:

• “Can a model classify ___ vs. ___ using ___ data?”
• “Does changing ___ improve model performance?”
• “Which feature matters more: ___ or ___?”

Examples (high school-friendly):

• Predicting recyclable vs. non-recyclable items from photos
• Classifying music moods from short audio clips
• Detecting spammy vs. normal messages (with a safe dataset)

2) Choose a dataset strategy

• Make your own dataset (best for controlled experiments)
• Use a public dataset (best for speed, but harder to claim originality)

3) Set a “minimum viable project” by day 7

By the end of the first week, your teen should have:

• A hypothesis
• 50–200 labeled examples (depending on project)
• A baseline model run (even if it’s bad)
• A place to record results (spreadsheet or notebook)

4) Use this judging checklist before printing the poster

• Do you have a clear independent variable?
• Did you keep other factors controlled?
• Can you explain one failure case?
• Do you show more than accuracy (precision/recall or confusion matrix)?
• Do you state a limitation and a next experiment?

If your teen wants a guided path, Intellect Council lessons can help them move from “cool idea” to a repeatable experiment: dataset, training, evaluation, and a confident explanation for judges. And if you want a strong starting point, use Jordan’s approach: start with the question, then earn the AI.