DART

DART is the best app for districts to generate bussing routes, bus stops based on accurate traffic, construction, and historical data. It ensures that everyone gets home safe and informed.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for DART

DART — The Story

One-liner

DART is a school transportation platform for route optimization, delay forecasting, and live tracking — built so districts can plan routes they can actually trust.

Built by Rahbe Abass + Aditya Mendiratta at HackKentucky as a proof of concept.

The problem: Kentucky's largest district can't get kids to school on time

We go to duPont Manual, a magnet school in Jefferson County Public Schools (JCPS). This isn't abstract for us — this is what happened to our own bus routes.

  • August 2023: first-day collapse. JCPS launched new routes and the system fell apart on day one. Buses were hours late, kids were stranded, and the district had to cancel 6–7 school days to recover.
  • $265,000 for routes drivers couldn't drive. The district had paid an outside consulting firm for the new plan. Drivers called the routes impossible — wrong turns, impossible timing, infeasible shifts.
  • $188M budget shortfall. Transportation cuts followed the financial crisis.
  • 13,000+ students lost their bus — including our school. In 2024, JCPS cut transportation for magnet and traditional schools, including duPont Manual.

That sequence is the core of why DART exists: an expensive, opaque, one-time consulting plan failed, there was no way to test it before kids depended on it, no way to see failure coming in the morning, and no way for families to know what was happening when it did.

Why bus routing is harder than it looks

It's tempting to think this is "just a maps app." It's not. School busing is a constrained, high-stakes routing problem:

  1. Every stop has a bell-time deadline, not just a destination. Being 10 minutes late isn't inefficient — it's a missed school day start.
  2. Every vehicle has a seat count and a wheelchair-lift limit. Capacity isn't one number. Accessibility constraints have to be honored on the right bus.
  3. Traffic and weather change the right answer daily. The fastest route on a clear Tuesday is not the right route on a rainy Monday.
  4. One infeasible route doesn't fail quietly. It strands real kids.

Legacy tools and static consulting reports optimize once, on paper, for the best case. Schools live in the worst case.

The answer: DART plans routes districts can trust

DART is one pipeline:

IMPORT → OPTIMIZE → COMPARE → SIMULATE → TRACK → NOTIFY

1. From CSV to route plan in minutes

Districts already have the data — it's stuck in spreadsheets. DART starts there:

  1. Upload roster, stop, driver, and vehicle CSVs
  2. Confirm column mapping and validate (e.g. home_addr → Student.address, wheelchair → Student.needs_lift, school_code → Student.school)
  3. Generate a route plan for any school

No SIS rip-and-replace, no months-long onboarding. Bring your messy CSVs, leave with a plan.

2. Optimization with real constraints (Google OR-Tools)

Underneath, DART solves a Capacitated Vehicle Routing Problem with Time Windows, fresh every morning:

  • Vehicles start at a depot and end at the school — every route is a real shift
  • Time windows come from bell schedules, not guesses
  • Vehicle capacity includes wheelchair-lift seats
  • Infeasible input returns structured reasons, not a bare "solver failed"

That last point matters. When the $265K consultant plan was undriveable, drivers knew it but the system didn't explain itself. DART tells you why a plan doesn't work so you can fix the input.

3. Fastest vs. Reliable — two plans, one click

Most routers give you one answer: shortest mileage. DART gives you two, side by side:

  • FASTEST (P50 median travel time): Minimum mileage and vehicles. Best case every day. Thin slack against the bell schedule.
  • RELIABILITY (P90 + late-risk classifier): Buffers the stops most likely to run late. Built to survive a rainy, high-traffic morning.

The Compare page shows mileage, vehicle count, and predicted on-time rate before a district commits. That tradeoff — cost vs. kids arriving on time — is currently invisible. We make it explicit.

4. Machine learning predicts delays before they happen

Gradient-boosted quantile models predict a median and a worst-case travel time for every segment, plus late-arrival risk.

Current prototype results (on synthetic data):

  • 91% lower travel-time error than a straight-line guess — 95 sec MAE vs. 1,064 sec
  • 98.7% F1 on the late-arrival classifier, 99.8% recall
  • 87.7% of real outcomes land under the P90 worst-case call

Honest labeling: trained on 17,600 synthetic segments, labeled synthetic in-product, with fallback to straight-line estimates if the model is unavailable. The architecture is real; the training data needs to become real district AVL/GPS history in production.

5. Monte Carlo simulation: rehearse a bad morning

Before rollout, DART runs a Monte Carlo digital twin — hundreds of simulated mornings with traffic and weather noise — so a dispatcher can see which routes break first. The consultant plan never got a dress rehearsal. Ours does.

6. While buses run: live GPS, dispatcher alerts, guardian ETAs

The same trip, three different views:

  • DRIVER: Starts a trip and posts GPS live, or taps Simulate Drive for a demo
  • DISPATCHER: Watches every trip live and gets alerted the moment one turns at-risk
  • GUARDIAN: Sees only their own kid's stop and ETA. No manifest, no other families.

Updates push live over WebSocket (Django Channels), with polling fallback on every client so a dropped socket doesn't mean a blind dispatcher.

Why this is necessary

  1. Kids lose learning time when buses fail. Canceled days and chronic lateness hit the students who can least afford it — those without a car fallback.
  2. Districts are flying blind. They buy a static plan, hope it works, and find out from angry phone calls that it doesn't.
  3. Drivers are set up to fail. Infeasible routes burn out drivers in the middle of a national driver shortage.
  4. Parents get no information. "Where is the bus?" shouldn't be a mystery in 2026.
  5. Budget cuts demand better math. When you're $188M short, you can't afford to waste mileage or to strand kids. You need to see the cost-reliability tradeoff explicitly.

Why DART is better than existing options

Typical alternatives today:

  • Legacy routing suites (e.g. Tyler/Versatrans, Transfinder, Edulog): Powerful but expensive, heavy to implement, desktop-era UX, and optimized for plan creation — not for daily delay forecasting, simulation, or family communication.
  • One-time consulting firms: Exactly what JCPS paid $265K for. A static PDF of routes, no live loop, no re-optimization, no accountability when drivers say it's undriveable.
  • Rideshare-style outsourcers / late-model startups (e.g. Zum, HopSkipDrive): Focused on providing drivers/vehicles or parent apps, not on giving the district itself an open, explainable optimization + forecasting engine.
  • Consumer maps (Google Maps / generic VRP demos): No bell-time windows, no fleet capacity, no wheelchair constraints, no school-day semantics.

DART wins on:

  • Reliability-first planning: P50 vs. P90 is a first-class choice, not an afterthought. We optimize for on-time arrival, not just short mileage.
  • Forecast + simulate before you commit: ML delay prediction and Monte Carlo twins mean you test the bad morning before kids live it.
  • Explainable feasibility: Structured infeasibility reasons instead of silent failures or "impossible" routes handed to drivers.
  • Minutes to first plan: CSV import with field mapping, not a 6-month IT project.
  • Closed live loop: Driver → dispatcher alert → guardian ETA in one system, with privacy by design (guardians see only their stop).
  • Built by the people it failed: We lost our bus. We know what the morning of chaos feels like. That urgency is a feature.
  • Cost: An open, lightweight pipeline vs. a $265K consulting engagement that can't be re-run tomorrow morning.

Honest status: proof of concept

DART is a proof of concept built in one hackathon weekend.

  • Students, GPS traces, and ML labels are synthetic
  • Not certified for real dispatch, navigation, or student-data compliance yet
  • No FERPA review, no real map-matching, no hardware AVL integration yet

We label synthetic data in-product and fall back gracefully when models are unavailable because trust starts with honesty.

What production needs next

  • SIS connectors (Infinite Campus, PowerSchool) to replace CSVs
  • FERPA review, access controls, audit logging for real student data
  • Hardware AVL / tablet GPS integration for drivers

Built With

Share this project:

Updates

Submission history