OnRoute — Real-Time Commute Intelligence for Delhi Commuters

Delhi has 4.5M daily DTC bus riders with zero reliable real-time information. Commuters like Priya leave home at 8:15 to catch the 8:22 bus — not knowing if it's early, late, or on time. Transit apps show outdated static routes. Google Maps has no live DTC data. The result: people add 10-15 min buffer time daily just to handle uncertainty. Multiply that by millions of commuters and it's a massive productivity loss built entirely on a data gap. Process Started by identifying the core user question: "Should I leave right now?" — everything was built around answering that one thing. First tried using real GTFS-RT feeds from DTC but no public API exists. Pivoted to a simulated live data pipeline with realistic Delhi route patterns, variance, and rush hour multipliers — statistically accurate even if not GPS-sourced. Built the prediction engine first: rush hour detection (7-9am, 5-8pm) multiplies base delay by 2.6x, confidence scoring uses route variance vs historical mean. Then wired the "leave-by" calculator on top: ETA minus walk time minus 2 min buffer. Added Groq + Llama 3.3 70B via a Vercel serverless proxy so the AI advisor works without exposing keys. Final layer was low-data and WhatsApp text mode — because real Delhi co Results Live app deployed at on-route-five.vercel.app with 5 Delhi DTC routes simulated in real-time. Prediction engine refreshes every 6 seconds with confidence scores between 40-95%. Leave-by calculator updates dynamically as walk time changes. AI advisor answers natural language queries using live bus context. Low-data mode cuts payload to under 1KB per update — functional on 2G. Demo scenario loads Priya's exact 8:22 bus situation instantly for evaluators to test without setup. Reflection I'd integrate real GTFS-RT feeds from DIMTS or DTC's internal systems instead of simulated data — the prediction logic is solid but real GPS positions would make confidence scores genuinely reliable. I'd also add push notifications so commuters get a "leave now" alert without opening the app, and build a WhatsApp bot that sends automatic updates to a user's number at their scheduled commute time. The architecture supports all of this — I'd prioritise real data sourcing as the next step.