Section 01 — Image Projects
Global image projects
These ten projects are designed to be universally relatable and exciting. They work for any student globally, use existing open datasets, and produce outcomes parents love to share. Each project is graded by complexity and includes exact classes to train.
🌿
Plant Doctor AI
Train an AI to diagnose sick plants by looking at their leaves. Healthy, wilting, or diseased — your model becomes the world's smallest plant clinic.
"My AI can tell if a plant needs water just by looking at it."
Classes to train
- Healthy leaf
- Wilting leaf
- Yellow / diseased leaf
Outcome
A printed 'Plant Prescription Pad' — the AI diagnosis + handwritten cure. Perfect photo with a real plant.
📱 "My 9-year-old just diagnosed a diseased plant with AI she trained herself. The plant is recovering."
🤟
Sign Language Translator
Teach an AI to recognise basic sign language letters or words. Build a bridge between the hearing and deaf communities using nothing but a webcam.
"My AI understands sign language."
Classes to train
- Sign for A, B, C
- Sign for Yes
- Sign for No
Outcome
Live demo video: kid signing and AI displaying the correct word in real time.
📱 "She taught herself sign language AND taught an AI to understand it. Grade 4."
🎨
Emotion Mirror
Train an AI to detect emotions from facial expressions. Happy, angry, surprised, confused — then use it to make an 'emotion mirror' people can look into.
"Stand here and the AI tells you what emotion it sees."
Classes to train
- Happy
- Surprised
- Thinking hard
- Confused
Outcome
A 'Mirror Station' poster + a results chart showing the AI's accuracy across different people.
📱 "She built an AI emotion detector and ran experiments on the whole family."
♻️
Smart Recycling Sorter
Train an AI to classify waste — plastic, paper, metal, organic. Then propose it as a real solution to your school's recycling problem.
"This AI could replace the recycling confusion at our school."
Classes to train
- Plastic
- Paper / cardboard
- Metal
- Organic / food waste
- E-waste
Outcome
A printed proposal document + demo video showing the AI sorting real items from a tray.
📱 "He built a working AI recycling sorter and is presenting it to the school principal."
🪨
Rock & Mineral Field Guide
Build an AI that identifies rocks and minerals from photos — just like the apps professionals use. Train it on samples you collect, then take it into the field.
"I trained an AI on 200 rock photos. Now it knows granite from basalt."
Classes to train
- Granite
- Basalt
- Limestone
- Quartz
- Sandstone
Outcome
A pocket field guide + an outdoor 'in the field' video testing the AI on actual rocks.
📱 "She built her own AI rock identifier and tested it in the field. It got 4 out of 5 right."
🏋️
AI Fitness Coach
Train an AI to recognise exercise postures — correct vs incorrect form. Build a coach that gives real-time feedback. Athletes and trainers actually use this kind of tech.
"Do a squat. My AI tells you if your form is right."
Classes to train
- Correct squat
- Incorrect squat (knees in)
- Push-up correct / incorrect
Outcome
Live demo video of the AI flagging incorrect form in real time + printed accuracy report.
📱 "My son built an AI personal trainer that checks your exercise form. He's 11."
🌍
Wildlife Species Monitor
Train an AI to identify local wildlife species from trail-camera style images — the exact same approach conservation scientists use to monitor endangered species.
"Wildlife researchers use this exact method to track tigers and leopards."
Classes to train
- Species A (e.g. crow)
- Species B (e.g. squirrel)
- No animal / false trigger
Outcome
A research report + presentation pitched as a real conservation tool with species accuracy table.
📱 "My 13-year-old produced a wildlife AI methodology report that reads like actual field research."
🩺
Medical Imaging Explorer
Using publicly available medical image datasets, explore how AI classifies scans — and critically examine where it fails, what biases it might have, and why that matters.
"AI reads X-rays in hospitals. I trained one to understand how it works — and fails."
Classes to train
- Normal scan
- Abnormal scan (educational dataset)
- Inconclusive / low confidence
Outcome
A printed ethics paper + slide deck. Portfolio material for secondary school applications.
📱 "My daughter's Grade 8 project explored AI bias in medical imaging. She wrote an ethics paper."
🌦️
Sky & Weather Observer
Photograph the sky every day for two weeks. Train an AI to predict weather from cloud types. Meteorologists actually use image classification for exactly this.
"I photographed the sky for 14 days and trained an AI to predict tomorrow's weather."
Classes to train
- Clear sky
- Cumulus (fair weather)
- Stratus (rain likely)
- Cumulonimbus (storm)
Outcome
A 14-day sky photo journal + accuracy chart comparing AI to a weather app.
📱 "She spent 14 days photographing the sky, trained an AI, and it matched the weather app 3/5 days."
🏛️
Cultural Heritage Classifier
Train an AI to identify traditional clothing, artefacts, or architecture from different cultures. Help preserve cultural knowledge in digital form.
"My AI can identify traditional textiles from 5 Indian states."
Classes to train
- Bengali textiles
- Rajasthani crafts
- Keralan architecture
- Punjabi attire
- Odishan art
Outcome
A digital museum page + interview recording with a family elder.
📱 "He built an AI to identify Indian textiles. Then interviewed his grandmother about the ones it got wrong."
Section 02 — Image Projects
India-rooted image projects
Projects built around things every Indian kid, parent, and dadi-nani will instantly recognise. From cricket to kolam to Bharatanatyam — these use the child's existing cultural knowledge as domain expertise.
🏏
Cricket Shot Classifier
Train an AI to recognise cricket batting shots — straight drive, pull shot, sweep, defensive block — just from a photo. Real sports analytics companies do this for the IPL.
"My AI watches cricket and names the shot. Just like a commentator."
Classes to train
- Straight drive
- Pull shot
- Sweep shot
- Defensive block
A live webcam demo where the AI names batting shots. Print the accuracy scorecard.
📱 "He trained an AI on cricket shots and it now names them faster than his grandfather."
🪔
Rangoli & Kolam Recogniser
Train an AI to tell apart rangoli, kolam, and mandala patterns — three ancient Indian art forms that look similar but have distinct rules.
"My AI knows the difference between a Tamil kolam and a Rajasthani rangoli."
Classes to train
- Rangoli (coloured powder, geometric)
- Kolam (rice flour, line-based)
- Mandala (circular, symmetric)
- Alpona (Bengal, white rice paste)
A 'Pattern Passport' booklet + live demo testing on new festival photos.
📱 "My daughter trained an AI to identify Indian folk art patterns. Sharing this Diwali."
🌾
Crop Health Monitor
70% of Indians depend on farming. Train an AI to detect crop diseases from leaf photos — the exact technology startups like DeHaat and Cropin use.
"Farmers lose crops because they can't always tell if a plant is sick. My AI can help."
Classes to train
- Healthy wheat / rice leaf
- Fungal infection (brown spots)
- Pest damage (holes)
- Drought stress (yellowing)
Printed 'Farmer's Field Report' + a mock WhatsApp card showing the AI diagnosis.
📱 "My son built a crop disease detector for Indian farmers. He's 12. The report reads like a scientist wrote it."
🛕
Temple Architecture Classifier
Train an AI to identify different styles of Indian temple architecture — Nagara, Dravidian, Vesara, and Hemadpanthi. Knowledge that takes archaeologists years to learn.
"Show my AI a temple photo. It tells you which style and which state it's from."
Classes to train
- Nagara (North India, curvilinear shikhara)
- Dravidian (South India, gopuram tower)
- Vesara / mixed (Deccan)
- Hemadpanthi (Maharashtra)
A 'Temple Spotter's Field Guide' + India map with classified temples marked.
📱 "She trained an AI on temple architecture and mapped every temple our family has visited. I'm framing this."
🥘
Dal & Sabzi Identifier
Every Indian household makes different dals and sabzis. Train an AI to identify them from photos — dal makhani vs chana masala vs palak paneer.
"My AI can tell dal tadka from dal makhani just by looking."
Classes to train
- Dal makhani
- Dal tadka
- Chana masala
- Palak paneer
- Aloo sabzi
A week's worth of meal photos + illustrated recipe cards + live tiffin-box demo at school.
📱 "She spent a week photographing home-cooked meals and tested the AI on her class's lunch boxes."
🌫️
Air Quality Sky Detector
India has 9 of the world's 10 most polluted cities. Train an AI to classify sky photos by pollution level — clear, moderate, hazardous — and build a 30-day AQI log.
"I photographed Delhi's sky every day for a month and trained an AI to measure pollution."
Classes to train
- Clear sky (AQI under 50)
- Moderate haze (AQI 100–200)
- Heavy smog (AQI 300+)
- Fog vs smog
A 30-day sky photo log + accuracy chart vs official AQI + research paper. Science fair gold.
📱 "My daughter compared her AI pollution detector to official CPCB data. The research paper is extraordinary."
🚌
Road Hazard Spotter
Train an AI to spot traffic hazards from dashcam-style photos — potholes, animals on road, overloaded vehicles. Real smart-city technology.
"Smart city systems use AI to monitor roads. I built a version in one afternoon."
Classes to train
- Pothole / road damage
- Animal on road
- Overloaded vehicle
- Wrong-side driving
A 'Road Safety Dashboard' poster + proposal addressed to the local municipal corporation.
📱 "His teacher submitted his road hazard AI to a Smart Cities competition. Grade 8."
🎭
Classical Dance Mudra Decoder
Indian classical dance has 100+ mudras — hand gestures that tell entire stories. Train an AI to recognise them. Every dance student has a superpower no one else does.
"I train in Bharatanatyam. Now I also trained an AI to read mudras."
Classes to train
- Pataka (open palm)
- Tripataka (three fingers up)
- Ardhachandra (half moon)
- Alapadma (lotus in bloom)
A printed 'Mudra Dictionary' + a 60-second live recognition demo video.
📱 "The demo video has been watched 400 times in our dance school WhatsApp group."
Section 03 — Useful Daily-Life Projects
Home & family image projects
These projects solve problems people actually have every day. The goal: a non-technical family member can open a link on their phone, use the AI, and get real value immediately. That's the "WhatsApp test."
🌿
Plant Thirst Detector
Half of all houseplants die from overwatering or underwatering. Train an AI to look at a plant and say: 'needs water now', 'doing fine', or 'overwatered — stop!'
"Our AI checks the plants every morning and tells Mum when to water them."
Classes to train
- Wilting / thirsty (drooping leaves)
- Healthy and firm
- Yellowing / overwatered
- Dry soil visible
A 'Plant Health Report' + recommended watering schedule derived from AI data.
📱 "She trained an AI to tell us when to water each plant. We haven't lost a single plant since."
📦 Data: PlantVillage — kaggle.com/datasets/vipoooool/new-plant-diseases-dataset
🚗
Car Needs a Wash Detector
Nobody agrees when the car is 'dirty enough' to wash. Train an AI to look at a car photo and classify: clean, mildly dusty, or needs urgent wash.
"Our AI decides when the car needs washing. No more arguments."
Classes to train
- Clean and shiny
- Light dust / fingerprints
- Visibly dirty / muddy
- Bird droppings / stains
A 7-day 'Car Cleanliness Log' + proof the AI predicted 'wash needed' correctly.
📱 "My son trained an AI to decide when our car needs washing. He's 11 and completely serious."
📦 Data: Self-collect 30 photos of family/neighbourhood cars. Use Google Images (Creative Commons) to supplement.
🍌
Fruit Freshness Checker
Train an AI to assess fruit freshness — fresh, nearly ripe, overripe, spoiled — from a photo. Point it at any fruit to know: eat today, tomorrow, or throw away.
"Point our AI at any fruit and it tells you: eat today, eat tomorrow, or throw away."
Classes to train
- Fresh / unripe
- Perfectly ripe — eat now
- Overripe — use in cooking
- Spoiled — throw out
A 'Weekly Fruit Report' — how many pieces were saved from being wasted + cost calculation.
📱 "He calculated we could save ₹800/month in wasted fruit using his AI. Grade 7."
📦 Data: Fruits-360 — kaggle.com/datasets/moltean/fruits | Fresh vs Rotten — kaggle.com/datasets/sriramr/fruits-fresh-and-rotten-for-classification
🏠
Room Cleanliness Inspector
Every parent asks 'have you cleaned your room?' Train an AI to inspect a room and classify: tidy, messy, or disaster zone. Let the AI be the judge — not Mum.
"My room gets checked by AI now. No more arguing with Mum."
Classes to train
- Tidy and organised
- Acceptable — a little cluttered
- Messy — needs work
- Disaster zone
A 30-day cleanliness log chart — a graph of room tidiness over a month.
📱 "We let our daughter's AI inspect her room every evening. The graph is the most honest thing in our house."
📦 Data: MIT Indoor Scenes — web.mit.edu/torralba/www/indoor.html | Self-collect 30 room photos in each state.
Section 04 — Study & School Projects
Study & school image projects
These projects use a webcam pointed at the student themselves — making the student both the builder and the subject. Parents immediately understand the value. These are also the easiest to deploy: just leave a browser tab open.
📚
Study Focus vs Distraction Detector
Train a webcam AI to classify: focused (looking at book/screen), distracted (looking away, phone in hand), or taking a break. Log every 5 minutes during study time.
"My webcam now knows if I'm actually studying or not. It's kind of scary."
Classes to train
- Focused — eyes on book
- Distracted — looking away
- Phone in hand
- Head down / possibly sleeping
A 'Study Efficiency Report' — pie chart of focused vs distracted time across one full week.
📱 "His report showed 58% focus time. He wasn't happy. He improved it to 74% the next week."
📦 Data: DAiSEE Student Engagement Dataset — iith.ac.in/~daisee-dataset (Indian-made) | Roboflow: universe.roboflow.com/search?q=class:drowsy
✋
Posture Alert
Indian students sit for 4–6 hours daily. Bad posture leads to real neck and back problems. Build an AI posture coach that detects slouching vs sitting straight — live on webcam.
"My AI yells at me when I slouch. My physiotherapist would love this."
Classes to train
- Good posture — straight spine
- Slouching forward
- Head tilting down
- Leaning to one side
A 'Posture Health Report' — chart of posture across a study session. Share with school health teacher.
📱 "She built a posture AI for herself and the school nurse asked for a copy of her report."
📦 Data: Self-collect is best — 150 photos of you + 3 family members in each posture = enough for 80%+ accuracy.
😴
Drowsiness Alert
Studying when sleepy is nearly useless. Train an AI to detect drowsiness from eye state and head position — and warn you before you lose the battle with sleep.
"The same AI used in trucks to prevent driver accidents — but for my study desk."
Classes to train
- Alert — eyes open, head upright
- Drowsy — heavy eyelids
- Yawning
- Eyes closed / head nodding
A 'Sleep vs Study' chart — time of day vs drowsiness detections. Best study hours recommendation.
📱 "He discovered he's most alert between 7–9pm. He restructured his whole study schedule. Grade 8."
📦 Data: MRL Eye Dataset — kaggle.com/datasets/prasadvpatil/mrl-dataset (84,000 eye images) | Driver Drowsiness Dataset — kaggle.com/datasets/ismailnasri20/driver-drowsiness-dataset-ddd
Section 05 — Business & Enterprise Projects
Business image projects
These projects are designed to create real, measurable value for small Indian businesses. A student who deploys one of these for a family business and produces a 'Business AI Report' is thinking like an entrepreneur — not a student.
🏪
Shopfront Cleanliness Inspector
Small shop owners can't always tell when their shopfront looks bad. Train an AI to inspect shop photos and flag: presentable, needs tidying, or urgent clean.
"We gave this to our uncle's shop. The AI checks the shopfront before opening every morning."
Classes to train
- Clean and presentable
- Dusty / cluttered
- Shutter dirty / paint peeling
- Entrance blocked / messy
A 2-week shop inspection log + business report connecting cleanliness to customer perception.
📱 "My daughter deployed her AI at her uncle's medical store. She presented a 2-week quality report. She's 13 and thinks like a consultant."
🥬
Sabziwala Produce Sorter
Every sabziwala manually sorts produce — fresh vs old, Grade A vs B. This takes hours. Train an AI to grade vegetables by quality from a photo in under a second.
"My AI can grade tomatoes the same way a trained mandi worker can — in under a second."
Classes to train
- Grade A — fresh, firm, vibrant
- Grade B — minor blemishes
- Grade C — soft spots, sell today
- Reject — spoiled
A live sorting demo video + business calculation showing time and money saved per day.
📱 "My son built a vegetable quality grader and tested it at our local sabziwala. The vendor asked him to come back. He's 14."
📦 Data: Vegetable Image Dataset — kaggle.com/datasets/misrakahmed/vegetable-image-dataset
🧾
Tiffin Quality Checker
India has 5 crore+ home tiffin providers. Biggest complaint: inconsistent quality. Train an AI to look at a plated dish and classify: excellent, acceptable, underfilled, or messy.
"My mum runs a tiffin service. Her AI now checks every box before it leaves the kitchen."
Classes to train
- Excellent presentation
- Acceptable — minor fix needed
- Underfilled — add more
- Spilled or messy — redo
A 7-day Quality Control Log. Shows how many boxes would have been sent below standard.
📱 "My daughter caught 6 under-filled boxes in one week. She saved us a customer complaint. Grade 8."
📦 Data: Indian Food Images — kaggle.com/datasets/l33tc0d3r/indian-food-images
🔧
Machine Servicing Detector
Small workshop owners miss early signs of machine wear. Train an AI to look at photos of equipment and flag: good condition, monitor closely, or service needed. Predictive maintenance is a ₹1000cr industry.
"My uncle owns a garage. His AI caught a worn fan belt before it snapped."
Classes to train
- Good condition
- Minor wear — monitor
- Visible rust / damage
- Critical — stop, fix immediately
A 'Preventive Maintenance Report' documenting issues flagged before they became expensive.
📱 "He spent a weekend at his uncle's garage and produced a maintenance report. His uncle is using it. Grade 8."
📦 Data: Surface Crack Detection — kaggle.com/datasets/arunrk7/surface-crack-detection | NEU Surface Defect — kaggle.com/datasets/kaustubhdikshit/neu-surface-defect-database
Section 06 — Audio Projects
Audio AI projects
Audio is the easiest modality for students to collect data — any phone mic works. Most projects need only 20 minutes of self-recorded sounds. Teachable Machine needs just 20 seconds per class to train. A working model can be built in a single class period.
Why audio projects are special: unlike images, you don't need external datasets for most of these. Students record the family coughing, the pressure cooker whistling, the dog barking — and that IS the training data. It's immediate, personal, and already inside the home.
🍳
Pressure Cooker Whistle Counter
Every Indian household uses a pressure cooker. 2 whistles for dal, 4 for rajma, 6 for mutton. Mum has to stand and count. Train an AI to count whistles automatically and alert when done.
"My AI listens to the pressure cooker and tells Mum when the dal is ready. She can go do other things."
Classes to train
- Pressure cooker whistle
- Background kitchen noise
- Silence / off
A working whistle counter running on phone in the kitchen — deployed and used by parents the same day.
📱 "My daughter trained an AI that counts pressure cooker whistles. We use it every single day."
📦 Data: 100% self-recorded. Record 30 whistle sounds and 30 background kitchen noise clips with any phone.
🤧
Cough Type Detector
Train an AI to classify cough sounds — dry cough, wet/productive cough, and normal throat clearing. Doctors actually use this distinction to assess respiratory health.
"My AI can tell a dry cough from a wet cough. Doctors actually use this distinction."
Classes to train
- Dry cough (sharp, tickling)
- Wet / productive cough (loose)
- Normal throat clearing
- Background noise
A 'Cough Diary' — 7 days of AI classifications for a family member with a cold, showing trend over days.
📱 "He ran a 7-day cough diary on himself during a cold. The wet cough peak matched the worst day of fever."
📦 Data: Coswara (IISc Bangalore) — coswara.iisc.ac.in — India-made cough dataset, free for research use.
🐕
Dog Bark Translator
Different barks mean different things — alert, hungry, playful. Train an AI to tell them apart and notify the right person for the right reason.
"My AI can tell the difference between our dog's alert bark and his hungry bark."
Classes to train
- Alert / intruder bark
- Playful bark
- Whine / hungry
- Silence
A 3-day 'Dog Bark Log' — every bark classified and timestamped. Families find this fascinating.
📱 "She trained an AI on our dog's barks. We now get notified on our phone when he barks at the gate."
📦 Data: ESC-50 includes dog bark category — github.com/karolpiczak/ESC-50 | Self-recorded is always better for this project.
😤
Snoring & Sleep Monitor
Heavy snoring is linked to sleep apnea. Train an AI to classify nighttime sounds: silence, light snoring, heavy snoring, and other disturbances. Help the family understand sleep quality.
"My AI listened to Dad sleep for a week. The report convinced him to see a doctor."
Classes to train
- Silence / normal breathing
- Light snoring
- Heavy / interrupted snoring
- Talking in sleep
A 'Weekly Sleep Report' showing % time in each category per night. A real health intervention tool.
📱 "His 7-night report for Dad showed heavy snoring 40% of the night. Dad made a doctor appointment."
📦 Data: ESC-50 breathing/snoring sounds. Self-recording nightly is most accurate.
🔔
Doorbell Alert for Elderly
Elderly grandparents or hard-of-hearing family members often miss the doorbell. Train an AI to detect these sounds and send a visual alert on a screen or tablet. Real assistive technology.
"My AI listens for our doorbell and flashes a big red alert on Dadi's tablet."
Classes to train
- Doorbell ringing
- Phone ringing
- Knock on door
- Someone calling 'Dadi / Nana'
- Background noise
A permanently running assistive device in the home. 1-week usage log + note from the elderly family member.
📱 "Dadi hasn't missed the door in 2 weeks. She keeps telling everyone. Grade 5."
📦 Data: ESC-50 includes doorbell, phone, knock. Record your specific doorbell too for best accuracy.
👶
Baby Cry Interpreter
New parents are exhausted trying to figure out why the baby is crying. Different cries — hungry, tired, pain — have distinct acoustic patterns. Train an AI to tell them apart in 3 seconds.
"My AI can tell the difference between a hungry cry and a tired cry — in under 3 seconds."
Classes to train
- Hungry (rhythmic, repetitive)
- Tired / sleepy (whiny)
- Pain (sudden, high-pitched)
- Discomfort / needing change
A 7-day 'Cry Log' with AI predictions vs what actually worked. Accuracy tracked per day.
📱 "My brother-in-law is begging for the link. Grade 7."
📦 Data: Donate-a-Cry corpus — github.com/gveres/donateacry-corpus (457 labelled baby cry recordings, free)
🔧
Machine Fault Detector by Sound
Mechanics can tell when a machine 'sounds wrong' — but only after years of experience. Train an AI to classify machine sounds: running normally, vibrating, bearing squeal, or about to fail.
"Experienced mechanics can hear a fault. My AI learned to do the same in one afternoon."
Classes to train
- Normal running sound
- Abnormal vibration / rattling
- Bearing squeal
- Knocking / misalignment
A 'Machine Health Log' for a real machine in a family business — with documented detections.
📱 "He caught an abnormal vibration before it became a breakdown. The mechanic was impressed. Grade 7."
📦 Data: UrbanSound8K — urbansounddataset.weebly.com | MIMII Machine Sound Dataset — kaggle.com/datasets/usmanyousaf/mimii-dataset-for-malfunctioning-industrial
🌧️
Rain Intensity Alert
Small businesses — chai stalls, open warehouses, farmers — need to know: light drizzle (stay open), moderate rain (cover up), heavy rain (close). Train an AI to classify rain sounds and send advance alerts.
"My AI hears the rain before you notice it. Our chai stall uncle closes up 10 minutes faster now."
Classes to train
- No rain / clear
- Light drizzle
- Moderate rain
- Heavy rain / storm
A 2-week rain alert log with cost saving calculation. Especially powerful during monsoon season.
📱 "She calculated we saved ₹1,400 in damaged stock because the AI warned us early. Grade 7."
📦 Data: ESC-50 rain category — github.com/karolpiczak/ESC-50 | Self-record during monsoon for best results.
🎵
Custom Voice Wake Word
Smart speakers cost ₹3,000–10,000. Build a free version using Teachable Machine that recognises YOUR custom voice command — lights on, play music, turn off fan — and triggers an action.
"I made my own Alexa using Teachable Machine. My wake word is my own name."
Classes to train
- Custom wake word (e.g. 'Hey ARIA')
- Command: lights on
- Command: play music
- Background noise
A live demo video: student speaks wake word → action happens in the room. The most dramatic demo of all.
📱 "Our house now responds to her voice. She's 10 years old."
📦 Data: Google Speech Commands — kaggle.com/datasets/neehakurelli/google-speech-commands (65,000 utterances)
🏪
Shop Ambient Noise Quality Monitor
The noise environment determines how comfortable customers feel and how long they stay. Train an AI to classify ambient sound quality and help shop owners improve the experience.
"Restaurants with the right noise level have 23% higher customer satisfaction. My AI measures this."
Classes to train
- Calm and pleasant (ideal)
- Busy but manageable
- Too loud / chaotic
- Too silent / uncomfortable
A professional 'Sound Environment Audit' report for a real business with recommendations.
📱 "The owner implemented his recommendation. The report was written by a 14-year-old."
📦 Data: UrbanSound8K restaurant/crowd sounds | ESC-50 indoor/outdoor ambient categories.
Section 07 — Resources
Free image datasets
All sources below are free and publicly available. Kaggle requires a free account. Roboflow Universe datasets can be downloaded as ZIP files directly. Download once and share via Google Drive with all students.
| Dataset | What it contains | Best for | Link |
|---|---|---|---|
| KagglePlantVillage (54,000 leaf images) | Healthy and diseased leaves across 38 plant categories. Gold standard for plant AI. | Plant Doctor, Crop Health, Plant Thirst | kaggle.com/datasets/vipoooool/new-plant-diseases-dataset |
| GitHub / IIT DelhiPlantDoc dataset | 2,569 images of 13 plant species with disease labels. Made by Indian researchers. | Plant Thirst, Crop Health | github.com/pratikkayal/PlantDoc-Dataset |
| KaggleFruits-360 (90,000+ images) | Fruit images across 131 fruit types. Best base for freshness projects. | Fruit Freshness, Sabziwala Sorter | kaggle.com/datasets/moltean/fruits |
| KaggleFresh and Rotten Fruits | Directly labelled fresh vs rotten. Perfect for freshness and tiffin quality projects. | Fruit Freshness, Tiffin Checker | kaggle.com/datasets/sriramr/... |
| KaggleMRL Eye Dataset (84,000 eyes) | Open and closed eye images under different lighting. Best for drowsiness detection. | Drowsiness Alert, Study Focus | kaggle.com/datasets/prasadvpatil/mrl-dataset |
| KaggleDriver Drowsiness Dataset (DDD) | Face + eye state images labelled awake vs drowsy. Direct use for study distraction. | Drowsiness Alert | kaggle.com/datasets/ismailnasri20/... |
| IITHDAiSEE Student Engagement | Student faces labelled by engagement: bored, confused, frustrated, engaged. Made in India. | Study Focus, Distraction Alert | iith.ac.in/~daisee-dataset |
| KaggleWaste Classification (25,000) | Organic vs recyclable waste images. | Smart Recycling Sorter | kaggle.com/datasets/techsash/waste-classification-data |
| KaggleGarbage Classification (12 classes) | 12 detailed waste categories: cardboard, glass, metal, plastic, paper. | Smart Recycling Sorter | kaggle.com/datasets/mostafaabla/garbage-classification |
| KaggleVegetable Image Dataset | 15 vegetable classes, cleanly labelled. | Sabziwala Sorter | kaggle.com/datasets/misrakahmed/vegetable-image-dataset |
| KaggleIndian Food Images | Indian dishes labelled by category — dal, biryani, roti, idli etc. | Tiffin Quality, Dal/Sabzi ID | kaggle.com/datasets/l33tc0d3r/indian-food-images |
| KaggleSurface Crack Detection | Concrete surface crack images — labelled cracked vs intact. | Machinery Servicing Detector | kaggle.com/datasets/arunrk7/surface-crack-detection |
| RoboflowRoboflow Universe (200,000+ datasets) | Largest collection of free, downloadable CV datasets. Search by class name. Export as ZIP. | Almost any project | universe.roboflow.com |
| GoogleOpen Images Dataset | 9 million images across 600 categories. Filter by class using the Open Images Explorer. | Any project needing diverse images | storage.googleapis.com/openimages/web/index.html |
Section 08 — Resources
Free audio datasets
| Dataset | What it contains | Best for | Link |
|---|---|---|---|
| IISc BangaloreCoswara — Indian cough dataset | 8,000+ cough, breathing, and voice recordings from Indian participants. Made in India during COVID research. | Cough Detector, Snoring Monitor | coswara.iisc.ac.in |
| GitHubESC-50 (2,000 environmental clips) | 2,000 five-second clips across 50 classes: rain, dog bark, doorbell, snoring, engine, baby cry, clock, fire. | Dog Bark, Doorbell, Rain, Snoring, Machine | github.com/karolpiczak/ESC-50 |
| WebUrbanSound8K (8,000 urban sounds) | 8,732 labelled clips: air conditioner, car horn, dog bark, drilling, engine, street music, jackhammer, siren. | Machine Fault, Rain Alert, Shop Ambient | urbansounddataset.weebly.com |
| GitHubDonate-a-Cry corpus | 457 baby cry recordings labelled by cause: hungry, tired, pain, discomfort, burping. | Baby Cry Interpreter | github.com/gveres/donateacry-corpus |
| KaggleGoogle Speech Commands (65,000) | 65,000 one-second utterances of 30 short words recorded by thousands of people. | Custom Wake Word, Voice Commands | kaggle.com/datasets/neehakurelli/google-speech-commands |
| KaggleMIMII Machine Sound Anomaly | Machine sounds in real factory conditions: pumps, fans, valves — normal vs anomalous. | Machine Fault Detector | kaggle.com/datasets/usmanyousaf/... |
| WebFreesound.org | Millions of freely licensed sound clips searchable by keyword. Find specific sounds you can't easily record. | Any project — search for specific sounds | freesound.org |
Section 09 — Deployment
Deploying your model as a real app
Getting a trained model onto a phone is the most important step — that's when it becomes genuinely useful. Here are all paths, ranked by ease. The "WhatsApp test" is the benchmark: if a non-technical family member can open a link on their phone and use the AI without explanation, you've succeeded.
First: export your model. In Teachable Machine, click "Export Model" after training. Always choose TensorFlow.js first — it gives you a hosted URL that works in PictoBlox, App Inventor, and any website without downloading anything.
The 3 export formats: (1) TensorFlow.js — hosted URL, works in any browser, best for most deployments. (2) TensorFlow Lite (.tflite) — compact file for Android apps, works offline. (3) Keras (.h5) — for Python projects on a computer.
Deployment paths by grade
Grades 3–5
PictoBlox desktop demo
Train on TM → get hosted TF.js URL → open PictoBlox → Machine Learning extension → paste URL → build block script: classify → say result. Demo on classroom laptop with live webcam. Best for: science fairs, parent days, classroom demos.
ai.thestempedia.com
Grades 6–7
MIT App Inventor → Android APK
Train on TM → get TF.js URL → go to appinventor.mit.edu → import TMIC extension (.aix file) → design UI with blocks → build → download .APK → install on Android phone → share APK via WhatsApp. Produces a real installable app.
appinventor.mit.edu · TMIC extension: community.appinventor.mit.edu/t/tmic/64411
Grade 8
Python + Streamlit web app
Export as Keras (.h5) → use the teachable-machine Python package (pip install teachable-machine) → build a Streamlit app in 10–20 lines → deploy to Streamlit Cloud (free) → permanent URL for anyone to use.
streamlit.io · pypi.org/project/teachable-machine
Note on PictoBlox: PictoBlox (made by STEMpedia in Ahmedabad) is the best starting point for Grades 4–7. It has a dedicated Machine Learning extension where you paste your TM URL directly and build logic with drag-and-drop blocks. Works on Windows and Mac. It stays on a laptop — not a phone — but it's the fastest path to a working, demo-ready project. For phone deployment, move to App Inventor.
iOS note: MIT App Inventor only produces Android APKs.
