📖 Chapter 6 of 12 · AI for New Learner- Senior 1 · Class 9
Chapter 6: Building Your First Real AI
In this chapter
- Aanvi from Nagpur — leading a hackathon team at age 14.
- The 6-step AI builder workflow used by every real engineer.
- Her actual build log: hour-by-hour from 0% to 91% accuracy.
- Where to publish your project — GitHub, Kaggle, IndiaAI.
- Mistakes that sink first-time AI builders (and how to avoid them).
The Big Idea
You have learnt enough to BUILD. Not just talk about AI — build it. With Python (Ch 1), maths (Ch 2-4), and logic (Ch 5), you have the full toolkit. This chapter shows the workflow real engineers use — the same 6 steps at Sarvam, ISRO, Niramai. Once you have shipped ONE project, you are no longer a student of AI. You are a builder.
Did you know?
Smart India Hackathon (SIH) is the world’s largest student hackathon — over 12 lakh students enter every year. Many top Indian AI engineers started as hackathon teenagers. Winners get cash prizes AND offers from ISRO, DRDO, and startups.
Builder Profile: Aanvi
Nagpur, Maharashtra · Aanvi, age 14
Aanvi formed a 3-person hackathon team at her Nagpur school. Their idea: an AI to spot crop diseases from leaf photos. Her uncle owns a farm in Vidarbha — every year he loses 15% of crops to disease spotted too late.
She and her friends entered the National Innovation Foundation hackathon. 36 hours. Three months of preparation. They had a working AI by the end — 91% accuracy on cotton-leaf images. They didn’t win. But the judges asked Aanvi for her code.
“What I learnt,” she says, “is that BUILDING is the entire game. Not having the best model — having ANY model that actually works on real data. Most adults never ship one. We shipped one in 36 hours.”
The 6-step workflow
Every real AI is built with the same six steps. (1) PROBLEM — pick one clear problem and write it in one sentence: “Spot diseased leaves from a photo.” (2) DATA — find or collect labelled data; 100 examples is the minimum, 1000 is good, 10,000 is excellent. (3) MODEL — pick the simplest model that could work; start with a decision tree, not a neural network. (4) TRAIN — split data 80/20, train on 80%, NEVER touch the 20% during training. (5) EVALUATE — measure on the 20%; accuracy alone is not enough, check precision, recall, fairness. (6) SHIP — save the model, write a 5-line README, push to GitHub, tell ONE person to try it.
WHY THIS MATTERS
EVERY Indian AI product you have heard of follows these 6 steps. Bhashini, Niramai, Sarvam — all began with one engineer doing exactly this. The biggest difference between you and them is YEARS of practice, not talent.
✎ BUILD LOG Crop Disease AI — 36 Hours
- Hr 0: Got the idea. Wrote one-line problem on whiteboard: “Spot diseased leaves.”
- Hr 4: Walked the field. 200 leaf photos, each labelled by farmers as healthy or sick.
- Hr 10: First model: 51% accuracy. Worse than a coin flip. Big disappointment.
- Hr 14: Spotted bug: training data shuffled wrong. Fixed. Accuracy → 73%. Real progress!
- Hr 20: Augmented images: rotate, flip, brightness. Accuracy → 86%.
- Hr 28: Two team members wanted to give up. I made chai. We continued.
- Hr 32: Final model: 91%. Wrote the 2-minute demo. Slept 30 minutes.
- Hr 36: Demo on stage. Got 2nd prize. Slept for 14 hours. Worth it.
✍ WORKED EXAMPLE Build an SMS Spam Filter
Problem: Build an AI that reads an SMS and decides ‘SPAM’ or ‘NOT SPAM’. Walk through the 6 steps with one decision at each.
- PROBLEM — ‘Given an SMS, predict if it is spam.’ Yes-or-no classification.
- DATA — use the UCI SMS Spam Collection (free, 5,574 labelled messages). Skip building your own.
- MODEL — start with Naive Bayes (Ch 4!). Works well for text, simple to debug.
- TRAIN — split 80/20; use scikit-learn’s CountVectorizer + MultinomialNB. About 5 lines.
- EVALUATE — accuracy 97%? That’s the WRONG metric. Spam is the minority class — check RECALL: did we catch enough real spam?
- SHIP — save the model to spam.pkl, push to GitHub, tell your family (most adults still get spam SMS).
Show answer
✓ Answer: A spam filter in ~200 lines of code. Truecaller uses the same approach in India.
Solved illustrations
1. What is the most common mistake first-time AI builders make?
Show answer
Answer: Skipping step 1. Without a clear, narrow problem (‘Spot diseased cotton leaves’), the project drifts. Spending 30 minutes writing the problem in one sentence saves 30 hours later.
2. Why is the 80/20 train-test split the rule?
Show answer
Answer: You need most data to TRAIN (~80%) and some to HONESTLY TEST (~20%). If you test on data you also trained on, you are cheating — and your model will fail in the real world.
3. How much data do I actually need?
Show answer
Answer: For a school project: 100 examples works. For something to deploy: 1,000+. For startup quality: 10,000+. Quality matters more than quantity — 200 well-labelled images beat 2,000 messy ones.
4. My model gets 99% accuracy. Should I celebrate?
Show answer
Answer: No — be SUSPICIOUS. 99% almost always means you tested on training data (cheating) or the data is imbalanced (99% of emails are NOT spam, so predicting ‘not spam’ always gets 99%). Check precision and recall.
5. Where do I publish my project?
Show answer
Answer: GitHub is mandatory — a public repo with README, model.pkl, and a 30-second demo. Kaggle for datasets. IndiaAI.gov.in showcases student projects. Streamlit hosts free web demos. Each adds a link to your CV.
Questions — Chapter 6
A. Builder workflow
- The 6 steps in order: Problem, Data, Model, Train, , Ship.
- The standard train-test split is / .
- The simplest text-classification model is Bayes.
- The world’s largest student hackathon is .
- Where every student MUST publish their AI project: .
B. Builder reasoning
- Why does Aanvi say ‘building is the entire game’?
- What goes wrong if you skip step 1 (Problem)?
- Why is 99% accuracy sometimes a RED flag?
- Pick one Indian AI from earlier chapters. Map it to the 6 steps.
- What is YOUR first project going to be? Write the one-line problem statement.
Ship-It Challenge
Open a GitHub account. Create a ‘my-first-ai’ repo. Add a README — even if empty.
Check your understanding — Chapter 6
Tap an answer to check it instantly. These are for practice — nothing is saved or sent.
Q1. What are the 6 steps of the AI builder workflow, in order?
- Ship, Train, Data, Model, Problem, Evaluate
- Problem, Data, Model, Train, Evaluate, Ship
- Model, Problem, Ship, Data, Train, Evaluate
Show answer
B. Problem, Data, Model, Train, Evaluate, Ship
Q2. What is the most common mistake first-time builders make?
- Using Python
- Skipping step 1 — not writing a clear, narrow problem, so the project drifts
- Publishing on GitHub
Show answer
B. Skipping step 1 — not writing a clear, narrow problem, so the project drifts
Q3. Why is the 80/20 train-test split the rule?
- It looks neat
- You train on ~80% and honestly test on ~20% — testing on data you trained on is cheating
- 20% is a lucky number
Show answer
B. You train on ~80% and honestly test on ~20% — testing on data you trained on is cheating
Q4. Your model gets 99% accuracy. Should you celebrate?
- Yes, immediately
- No — be suspicious: you may have tested on training data, or the data is imbalanced
- Yes, 99% is always perfect
Show answer
B. No — be suspicious: you may have tested on training data, or the data is imbalanced