MODULE 08 · 2+ HOURS · CAPSTONE

Build Your
AI + IoT Mini Project

Plan a small AI + gadget project step by step — pick a problem, build, test, demo, and say what you would improve.

Pick a problem
Build & test
10 Quiz Questions
8 Real-World Examples
10 Quiz Questions
🎯 Scope 🧰 Stack 🔌 Integration 🎤 Demo day
↓ scroll to begin

Start with a clear problem

A good student project answers four questions on one page: Who has the problem? How will we know if we solved it? What do we build this month? What do we not build?

💡 This module ties Modules 4–7 together: sensors (4), IoT messages (5), optional vision or simple AI (6), and honest testing plus ethics (7).

Start with a story, not a technology. Weak pitch: “We use MQTT and TensorFlow.” Strong pitch: “The school garden dries out on weekends — we alert the club when soil is too dry.”

One-page project brief template:

  1. User — who has the problem? (gardener, lab tech, student, teacher)
  2. Problem — what hurts today? (overflow, wasted heat, late bus)
  3. One metric — how we know it worked (alert within 2 min; 90% correct labels)
  4. Baseline — what they do now without your system
  5. In scope — sensor + chart + one alert (be specific)
  6. Out of scope — mobile app store, payments, social network, etc.
SCOPE — WHAT YOU WILL AND WILL NOT BUILD IN SCOPE sensor + Wi‑Fi + chart + alert OUT OF SCOPE app store, payments, full social site

Figure — Saying “no” to extras is a sign of a mature project plan.

Project idea types — what each needs

Idea typeTypical sensorsTypical outputExample metric
Environment monitorTemp, humidity, CO₂Web chart, email if badAlert within 2 minutes
Fill-level / distanceUltrasonic, IR“Bin 80% full” on phoneLevel correct ±10%
Security / accessDoor, motion (PIR)Log + buzzerZero missed opens in test week
Plant / food qualityCamera (vision)OK / not OK label90% on 50 test photos
Energy saverLight, motion, smart plugTurn off lights when emptykWh saved vs baseline week

Good vs weak project goals

Weak goalStronger goal
“Use AI”“Alert when soil moisture below X for 30 minutes”
“Make an app”“Show live temperature on a web page updated every minute”
“100% accurate”“Correct on 45 of 50 test readings”
“Revolutionise farming”“Help one school greenhouse avoid dry weekends”
Small is good One sensor, one chart, one alert beats a huge plan you cannot finish.
Demo beats slides Live reading from a real sensor — even if ugly — impresses more than mockups.
Honest limits Say what failed. Teachers trust “here is what broke” more than perfect claims.
Example — compost bin: Ultrasonic sensor on lid measures distance to food waste. ESP32 sends reading every 5 minutes. Web page shows fill %. If above 80% for one hour, send email to garden club. Out of scope: native phone app, smell sensor, automatic collection truck.
Example — quiet study room: CO₂ sensor in library corner. Chart turns orange above 1000 ppm. Sign says “air stuffy — open window.” Metric: students see CO₂ drop within 5 minutes of opening window in a test.
Try it · Write six lines for your project brief

User, problem, metric, baseline, in scope, out of scope. Swap with a partner — can they explain your project without jargon?


Sensor → internet → brain → action

Every project is a chain: something in the real world is measured, a message travels over the internet, software decides, and something happens on a screen or a buzzer.

Think in messages, not magic. Your board does not “know” the temperature — it reads a number and sends a small text packet. The server stores it. A rule or tiny model says “if hot, alert.”

Sample message (JSON is fine):

{"device": "greenhouse-1", "temp_c": 22.5, "humidity": 61, "time": "2026-05-16T10:00:00Z"}

Typical student stack: sensor board → home Wi‑Fi → small server or cloud function → simple rule or small vision model → web page, LED, or email.

STUDENT PROJECT PIPELINE sensor Wi‑Fi server rule orsmall AI LED/page

Figure — Each box is something you can demo and explain in one sentence.

ONE READING TRAVELS board reads sends JSON server saves chart updates If Wi‑Fi fails, decide: queue on board, show “offline”, or retry

Figure — Draw this for your poster — judges understand arrows faster than code.

Parts of the stack — what each piece does

PartJobUsed forExample
SensorMeasure the real worldTemp, distance, light, motionDHT22 humidity
BoardRead sensor, send dataConnect sensor to Wi‑FiESP32, Arduino + shield
InternetCarry messagesHome or school Wi‑FiRouter in lab
Server / cloudStore data, run logicCharts, alerts, logsSmall Python app online
BrainRule or tiny modelDecide when to alertIf temp > 30 °C, email
OutputShow or actHuman sees resultWeb chart, buzzer, LED

Board choices — what students often pick

BoardGood atWatch out for
ESP32Wi‑Fi built in, cheap, many tutorialsPower use; secure your Wi‑Fi password
Arduino UnoSimple sensors, huge communityNeeds extra part for Wi‑Fi
Raspberry PiRuns full Linux, camera projectsMore setup; treat it like a small PC
Rules first “If distance < 10 cm for 5 minutes” is a valid brain — you do not need deep learning for every project.
Draw before you code A box diagram saves hours when the Wi‑Fi stops working.
Name your devices greenhouse-1 beats sensor_abc so logs make sense.
Example — hallway heat: ESP32 + temperature sensor in corridor. Every minute sends JSON to school server. Web page shows line chart. Rule: if above 28 °C for 15 minutes, email caretaker. Vision not needed.
Example — lost-item camera: Raspberry Pi camera takes photo every hour. Small image model (Module 6) labels “backpack present / missing” on a shelf. Chart for drama club props room. Ethics: only film the shelf, not people walking by.
Try it · Sketch your pipeline on paper

Five boxes, four arrows, one sample JSON message. Show a classmate — can they predict what happens when the sensor breaks?


A demo that works once is not enough

Real engineering means trying to break your own work — unplug Wi‑Fi, send nonsense numbers, ask a friend to run it. Write down what failed; that is how you improve.

⚠️ Leakage cheat (from Module 3): If you train on data that already contains the answer (e.g. repair logs that include the repair date), scores look amazing and fail in real life. Keep a test set the model never saw during training.

Three layers of testing:

  1. Unit — does one function work? (convert reading to percent full)
  2. Integration — does sensor + Wi‑Fi + server work together?
  3. User test — can someone else follow your README and see the demo?

Document failures on purpose. A slide titled “What broke” shows maturity. Examples: “Wi‑Fi dropped at 3 pm; chart froze until refresh.”

TEST → BREAK → FIX → REPEAT test break fix repeat

Figure — Same loop professionals use before a science fair or industry demo.

Testing checklist — try to break your own demo

TestWhat you doWhat good looks like
Unplug Wi‑FiPull router or use airplane mode on boardClear “offline” message, not fake numbers
Wrong sensor valueSend -999, empty, or huge numberIgnores or warns; server does not crash
Two people run itFriend follows README onlyThey repeat demo without you touching keyboard
Held-back photosVision: 20 new pictures never used in trainingHonest accuracy count on slide
Honest limits slideList failures and weak spotsAudience trusts your numbers
Ethics checkWho could be harmed or watched?Camera points only where school allows

Common failure modes — and what to do

What went wrongLikely causeFix idea
Chart flatlinesWi‑Fi or wrong URLLog last error on OLED or serial monitor
Random spikesLoose wire, bad powerAverage last 5 readings
Model always says “OK”Training only easy photosAdd hard examples; retest
Email never arrivesSpam folder, API keyTest with your own address first
Demo works only on your laptopHard-coded pathsREADME with exact steps

Simple four-week timeline (team of 2–3)

WeekGoalDeliverable
1Scope + one sensor reading on serialProject brief, photo of breadboard
2Wi‑Fi + server + live chartURL that updates every minute
3Rule or model + alertOne real alert you triggered on purpose
4Break tests + poster + talk“What failed” slide, model card (Module 7)
Example — compost team: Week 3 email alert fired twice during lunch test. Week 4 unplugged Wi‑Fi — chart showed “last seen 10:04, offline” instead of lying about fill level.
Example — fruit sorter: Model scored 94% on training photos but 71% on new bruised apples from the market. Slide said: “Needs more bruised examples — not ready for factory.”
Checklist · Before presentation day

Present results clearly

Judges and classmates remember stories, not buzzwords. Walk them through the problem, live demo, honest numbers, limits, and what you would do next.

Five-minute talk outline (about one slide each):

  1. Problem — who suffers and how today?
  2. Your system — box diagram (Topic 02)
  3. Live demo — real sensor or real photo through model
  4. Numbers — one metric from your brief; how you measured it
  5. Limits + next step — what failed; two weeks more work

Poster tips: big diagram in the centre, QR code to live chart, short bullets, no tiny paragraphs. Photos of hardware beat screenshots of code.

GOOD FINAL TALK ORDER 1 problem 2 demo 3 numbers 4 limits 5 next step

Figure — Tell a short story, not a jargon list.

What judges often look for (plain English)

AreaStrong projectWeak project
ProblemClear user and pain“We used AI because it is cool”
DemoLive or recorded real runOnly mockups
HonestyAdmits limits and failuresClaims 100% accuracy
EthicsCamera/consent thought throughFilms people without saying so
TeamworkEveryone can explain one boxOne person does all talking

Slide titles that work

Instead of…Try…
“Deep learning pipeline”“How we know the bin is full”
“Results”“45 of 50 test readings correct”
“Conclusion”“What we would fix with two more weeks”
“IoT architecture”“From sensor to text alert in 2 minutes”
Rehearse the demo twice — once with good Wi‑Fi, once with phone hotspot only.
Bring the thing — board on table beats slides-only.
End with thanks + questions — leave 30 seconds for one good question.
Example — greenhouse team talk: “Weekends the soil dried out. We built a $40 sensor that texts the club president. In tests, 9 of 10 dry spells triggered within 2 minutes. Wi‑Fi in the shed is weak — next step is a data logger on the board.”
Example — poster layout: Centre: pipeline diagram. Left: problem photo. Right: chart screenshot. Bottom: QR to live page + three-line model card from Module 7.
Try it · Practice your first 30 seconds out loud

Who is the user, what is wrong today, what you built — no acronyms. Time yourself. If it takes more than 30 seconds, shorten.


Quick Knowledge Check

10 easy questions on planning and showing a student project. Instant feedback on every answer.

Score: 0 / 0

Key Takeaways

Module 8 in short: clear goals, honest testing, and a simple story beat a flashy demo that only works once.

↑ Back to Top
📚 Project references:
• Git documentation — git-scm.com/doc
• FastAPI docs (APIs) — fastapi.tiangolo.com
• Edge Impulse (embedded ML) — edgeimpulse.com