AI science fair projects built around a question, not a demo.
The most common reason a strong AI build does badly at a science fair is that it is a product, and a fair is judging science. Same code, different framing, completely different result.
Classes that cover the model and the method
A fair project needs data handling and experimental discipline as much as it needs code.
Python and AI classes for kids and teens
The usual starting point. Language fundamentals to the level where a first real project becomes possible.
View course
Build machine learning models in Python
For students who already code and want a model they trained themselves, not a tutorial they copied.
View course
Data science course
Collecting, cleaning and interrogating data, which is most of the work in any real project.
View course
AI, ML, Python and Java track
The longer route for students aiming at research-grade work rather than a single app.
View courseIn short
An AI science fair project is judged as science, which means it needs a question, a method, controls and a result, not simply working software. Regeneron ISEF is not open entry: finalists qualify by winning a Society-affiliated fair, and projects may include no more than 12 months of continuous research. Teams are capped at three members.
Modern Age Coders teaches students to build the model and, just as importantly, to design the experiment around it. To discuss a fair project, use the form below or message us on WhatsApp at +91 91233 66161.
The short version
- A fair judges a question and a method. "I built an app" is not a question.
- ISEF is qualified into, not entered. Finalists come through 365 affiliated fairs across more than 60 countries.
- The research window is capped at 12 months of continuous research, so an early start does not mean an earlier start is always better.
- Team projects are limited to three members.
- The strongest school AI fair projects use data the student collected, and state clearly what the result does not prove.
The rules people find out about too late
Every one of these has ended somebody's season. All are from the official rules.
ISEF is not open entry
Students in grades 9 to 12 must compete in and win through a Society-affiliated fair to earn a place. Finalists are drawn from 365 affiliated fairs in more than 60 countries and territories.
Source: Society for Science, ISEFA hard cap on continuous research
Each project may include no more than 12 months of continuous research. Work stretching over several years has to be scoped and presented carefully against this rule.
Source: Society for Science, rules for all projectsTeam size is capped
Team projects may have no more than three members. Larger groups have to split the work into genuinely separate projects.
Source: Society for Science, ISEF rulesIt is genuinely international
The fair draws roughly 1,700 finalists, with a substantial majority of participating countries outside the United States, so students abroad have a real path in through their national or regional affiliate.
Source: Society for Science, ISEF FAQFourteen AI fair projects with a real hypothesis
Each one is written as a question that can be answered wrongly. That is the test of whether it belongs at a fair.
Model behaviour questions
Cheap to run, strong scienceDoes more training data help less than better labels?
Hold the model fixed. Compare a large noisy dataset against a smaller carefully labelled one, and measure where each wins.
A clean two-variable experiment with a genuinely uncertain answer. Judges like uncertainty.
How much can you shrink a model before it breaks?
Progressively reduce a model and plot accuracy against size, identifying the point where performance collapses rather than declines.
Produces a curve, and a curve with a knee in it is a finding.
Does image augmentation help on your own photographs?
Test whether standard augmentation techniques improve results on a dataset you collected, where the assumptions may not hold.
Testing a standard technique on non-standard data is a legitimate small research question.
Which errors does the model make that a human never would?
Classify the failure cases and test whether human annotators make the same mistakes on the same inputs.
Comparing to human performance gives you a control group, which fairs reward.
Does the model rely on the thing you think it does?
Occlude or remove parts of the input systematically and measure what actually drives the prediction.
Often produces a surprising answer, and surprise is what makes a judge remember a board.
Applied questions with local data
You collect the dataCan a phone camera classify local water quality indicators?
Collect samples with known measurements, photograph under controlled lighting, and test how well colour alone predicts the reading.
Local, testable, and the controls for lighting are exactly the rigour judges look for.
Does air quality near your school follow the pattern the public sensor claims?
Deploy your own sensors, compare against the nearest official station, and quantify the disagreement.
Comparing against an authoritative source is a proper method with a clear result.
Can leaf photographs predict disease earlier than visible symptoms?
Photograph the same plants over weeks, label by eventual outcome, and test whether the model detects it before a person can.
The time dimension gives you a genuine claim: earlier than the human baseline, or not.
Does a model trained on one accent work on yours?
Test a speech model against recordings in your local accent and quantify the performance gap.
A measurable, socially meaningful gap that you are well placed to document.
Can sound alone identify local bird species?
Record in your area, label with expert help, and report which species are separable and which are not.
Ecology fields love this, and the confusions between similar species are the interesting result.
Method and fairness questions
Design the measurementDoes the order of examples change what a language model answers?
Hold the content fixed, vary the order, and measure how much the answer moves.
A simple, rigorous test of a property people assume is stable.
Can a model detect its own uncertainty usefully?
Measure whether a model's confidence score actually predicts whether it is right, across many cases.
Calibration is a real research area and the experiment is within reach at school level.
Does a fairness fix reduce bias or move it?
Apply a published mitigation technique and measure whether the disparity shrinks overall or shifts to a different group.
The honest answer is often uncomfortable, and honest uncomfortable answers score well.
How little labelled data do you actually need?
Plot performance against labelled examples to find the point of diminishing returns for your specific task.
Practical, quantitative, and the resulting curve is a genuine contribution for that task.
What a fair board needs that an app demo does not
The framing differences that decide how a judge scores the same underlying work.
A falsifiable question
Written so that a specific result would prove you wrong. If no outcome could disprove it, it is not a hypothesis.
A control or baseline
What you are comparing against. Without a comparison there is no result, only a number.
A documented method
Detailed enough that another student could repeat it and get the same answer.
Data you can account for
Where every data point came from, how many you discarded, and why. Judges ask this.
Statistics, not vibes
Repeat runs, variation reported, and a clear statement of whether the difference you found is meaningful.
Stated limitations
What your result does not prove. Volunteering this is one of the strongest signals available to a school student.
A season, working backwards
Fair dates vary by region and affiliate, so confirm your local fair's calendar first and count back from there.
Find your affiliated fair
Locate the Society-affiliated fair that serves your region and read its rules and dates. This decides everything downstream.
Write the question and the method
Before any code. If the question cannot be answered wrongly, rewrite it.
Collect data and build the pipeline
Usually the longest phase, and the one students underestimate most.
Run the experiment properly
Repeat runs, record variation, and resist changing the question to match the result.
Write it up and build the board
Method, results, limitations. The board is a summary of the work, not the work itself.
Rehearse being questioned
Judges interview you. Practise answering "how do you know?" about every claim on the board.
Fair seasons repeat, so early starts compound
A fair is annual. A student who begins young enters several times, and each entry is visibly better than the last.
Starting at 6 or 7, building for years
Children who start early get to fail at a fair at ten and come back at fourteen knowing what a control is. That progression is far more convincing than one polished entry.
- Ages 6 to 9Block coding and logic. Finishing small things, which is the habit everything later depends on.
- Ages 10 to 13Real code in Python. First projects that another person can actually run.
- Ages 14 to 16AI and machine learning, own datasets, work that goes in front of real users.
- Ages 17 to 18Portfolio, competitions and research-grade work, with years of history behind it.
By the time it matters, the repository is years deep. That cannot be assembled in a hurry, and it is obvious to anyone who looks.
A deadline in weeks, not years
A first entry with a genuine question and a clean method can be built in a season. Note the research window is capped at twelve months, so more time is not automatically more project.
- Weeks 1 to 2Scope honestly against the time left. Most ideas get cut down here, and that is the useful part.
- Weeks 3 to 6Build the smallest version that genuinely works, with the repository open from day one.
- Weeks 7 to 8One real user, then fix what tripped them up.
- Final weekREADME, recorded demo, and rehearsal for being questioned about it.
Smaller and finished beats ambitious and broken. We will tell you before you start if the deadline does not allow honest work.
Bring us the fair date first. Everything else, including whether the idea is the right size, is decided by how many weeks are actually left.
We teach the experiment, not just the model
Most coding classes stop when the code runs. A fair project is only starting at that point.
The question comes before the code
Our mentors work with the student on the question first, then the build, then the analysis. The insistence on a baseline, repeat runs and a written limitation is where a fair project separates from a school assignment.
- A hypothesis written so that some result would prove it wrong
- A method recorded before the data arrives, not after
- Controls and repeat runs, because one run is an anecdote
- Results written up with the limitations stated plainly
- Rehearsal for judging, where every claim gets a "how do you know?"
Students usually need this much before a fair project is realistic.
Python and AI classes for kids and teens
Beginner
Build machine learning models in Python
Intermediate
AI, ML, Python and Java track
Advanced
What it costs to prepare
Fair mentoring runs inside normal classes at the normal fee. We do not charge a competition premium.
Group class
Batches of 5–8. Good for building the fundamentals.
₹1,499 per monthMini batch
3–4 students. More airtime per student.
₹2,999 per monthOne to one
1:1. The usual choice for a project with a deadline.
₹7,500 per monthPrices are per month and shown per region. Fair entry fees, materials and travel are set by the fair, not by us, and are never paid to us. See full pricing for every subject and tier.
What our students build
Real projects from Modern Age Coders students on Student Labs.
GuardianX, an AI safety layer for the internet
A student build aimed at filtering harmful content. The interesting engineering is in the false positives, and the student can explain them.
HairSync, an AI hairstyle try-on
Image manipulation where the output is judged by eye, so the student had to define what good even meant.
Life Span Tracker
Date arithmetic, which is one of the classic places beginner code quietly goes wrong.
SkyCast, a weather forecast app
Third-party API, error states, and what to show when the network fails. Unglamorous, and exactly what reviewers probe.
Tell us about the fair and the deadline
Tell us which fair and which date. We reply within one business day, and if the deadline is too close for honest work we will say so.
Deadline already close?
Message us with the date. Scoping a fair project down to fit the time left is a conversation worth having early.
Your details are used only to reply to your enquiry and are never shared or sold. We mentor students through their own work and never produce a project for them.
Frequently asked
How does a student qualify for ISEF?
By competing in and winning through a Society-affiliated local, regional, state or national fair. Finalists come from 365 affiliated fairs across more than 60 countries. There is no route that skips the affiliated fair.
Can international students compete at ISEF?
Yes. The affiliate network spans more than 60 countries and territories, and a large share of finalists come from outside the United States. Find your national or regional affiliate first.
How long can the research take?
A project may include no more than 12 months of continuous research. Longer programmes of work need to be scoped so that the entered project falls inside that window.
Can students work in teams?
Yes, up to three members per team project.
What makes an AI project suitable for a science fair rather than a hackathon?
A fair judges a question, a method, controls and a result. A hackathon judges a working build. The same model can serve both, but the fair version needs a hypothesis that could have been proved wrong.
Does the project need to be original research?
It needs a question you answered yourself with a method you can defend. Testing an established technique on data you collected in your own area is legitimate and often stronger than an over-ambitious novel claim.
What if the hypothesis turns out to be wrong?
That is a result. Judges respond well to a clean experiment with a negative result and an honest discussion. Changing the question after seeing the data is the thing to avoid.
Can Modern Age Coders help with a fair project?
We teach the coding and the experimental design, live and with a mentor. We do not do the project for the student, and a project a student cannot explain would defeat the purpose.