Problem → Solution → Tech → Conclusion
Emmanuel · Samira · Kevin · Kenny · Hassan · CMU-Africa Bridge Program 2026
How long did you wait for a bus this week? A Kigali commuter leaves work at 6 PM and reaches home after 9 PM, and Rwanda’s transport regulator has said the average wait can reach one hour.[1]
Buses wait for passengers while passengers wait for buses, because neither side can see the other. Off-peak scheduled buses left with only 10–15 passengers,[2] while at peak and holiday times travellers waited several hours.[4] Furthermore, Kigali needed about 500 buses but had 229 in 2022.[1] Nobody measures who is waiting, where, or for which destination.
Kigali’s buses cannot see where passengers are waiting, so they often run either too empty or too full.
In December 2025 the city launched “Generation 2”, with scheduled departures and real-time bus tracking.[3] However, a schedule is planned in advance, while demand changes by hour, stop and destination.
During the scheduled-service pilot, off-peak buses left with only 10–15 passengers, and commuters asked for schedules adjusted to demand.[2] In contrast, during the 2025 holiday peak, demand exceeded bus capacity and travellers waited several hours.[4] Furthermore, research in Lagos found that demand-based scheduling could cut average waiting time by up to 80% in simulation.[6]
Therefore, Kigali needs more than extra buses; it needs to know who is waiting, so every bus leaves for a reason.
Every bus stop becomes a sensor. Every bus leaves for a reason.
The kiosk counts who is waiting and where they are going.
HQ gets a live “release a bus” recommendation and approves it.
Passengers get a fair queue number and a live ETA.
Tap&Select works with Generation 2: schedules plus real demand.[3]
A fair place in line and a live ETA, with no smartphone needed. They choose their stop hands-free, by voice or a swipe in the air, and hear their ticket number read out loud.
The aim: fuller buses, fewer empty off-peak trips, less fuel wasted.
Live demand per stop and destination, to plan routes and fleets with data instead of guesswork. Public transport is about 20% of daily trips, and the city aims to reach 50%.[5]
Tap&Go; simulated reader in the prototype
Voice or a swipe over the sensor
On-device, no internet
Tap Confirm on screen; ticket e.g. R2-014
Kiosk shows the live ETA
| Feature | Benefit | Status |
|---|---|---|
| 1Live demand → dispatch loopTickets reach HQ within seconds; rules recommend a release; a person approvesBuses leave for a reason; HQ stays in control | Buses leave for a reason; HQ stays in control | Real |
| 2Fair first-come queueNumbered tickets, one per card; the oldest board first, up to free seatsBoarding follows ticket order, and places can’t be held | Boarding follows ticket order, and places can’t be held | Real |
| 3Live bus trackingDriver's phone GPS: stop-by-stop progress, arrival within 50 mPassengers see where their bus is | Passengers see where their bus is | Real |
| 4Swipe to choose a stopSensor direction engine, no ML; 85–95% per direction in 20-try testsChoose a stop without touching the screen | Choose a stop without touching the screen | Real, rule-based |
| 5On-device voice keywordsEdge Impulse model on the Arduino; accuracy not yet measuredRecognition needs no internet; audio stays on the board | Recognition needs no internet; audio stays on the board | Real |
| 6Accounts and rolesFive roles; rights enforced in the databaseEach person sees and does only their part | Each person sees and does only their part | Real |
| 7Access for everyoneEnglish + Kinyarwanda (draft), spoken numbers, tap-a-stop fallback, /liveLess reading needed, and no smartphone | Less reading needed, and no smartphone | Real |
| 8Seats and why people leaveSeats counted from tickets; cancel reasons so far from our tests onlyShows full buses and why people give up | Shows full buses and why people give up | Estimated |
Edge Impulse model · PDM microphone, 16 kHz, 1 s windows
| Label | Selects |
|---|---|
| remera | R2Remera |
| nyabugogo | R1Nyabugogo |
| kimironko | R3Kimironko |
| downtown | R4Downtown |
| noise | Nothing (rejected) |
| unknown | Nothing (rejected) |
Selects when one stop scores at least 0.80 twice in a row (one result every 250 ms). Exported 2026-10-28.
APDS9960 built-in direction engine (no ML) · read on its own thread
| Over the sensor | Selects |
|---|---|
| Swipe up | R1Nyabugogo |
| Swipe left | R2Remera |
| Swipe down | R3Kimironko |
| Swipe right | R4Downtown |
A hand passing over the sensor within about 10 cm is read as left, right, up or down. The earlier tap-count rules are still in the firmware as a fallback build, and the data logger for a trained Edge Impulse tap model is ready.
One kiosk, four destinations, live dashboard
2–3 stops on one route
Every stop a demand sensor
Today, people wait for buses. Tomorrow, buses will know who is waiting.