RajivKumar_OpsLead OP
Operations Manager · Bangalore
2 days ago
We've been running Google OR-Tools combined with a custom-trained LSTM model for traffic prediction on our Bengaluru delivery network since late 2025 and the results have been genuinely surprising — we cut average delivery time per route by around 23% without adding a single vehicle. The real secret nobody talks about is that the AI alone isn't enough; you have to feed it hyperlocal data like apartment gate timings, festival calendars, and rain-season road closures. We built a small pipeline using Python that pulls live data from the city's open traffic API and injects it as dynamic constraints into the solver every 90 minutes. Check out developers.google.com/optimization if you haven't started — the vehicle routing section is extremely well documented and free to use.

SarahMontgomery_SCM
Supply Chain Analyst · London
1 day ago
Worth mentioning that OpenStreetMap's Valhalla routing engine is a seriously underrated open-source alternative if you're not ready to commit to a commercial API budget. We integrated it with a reinforcement learning agent trained on our own historical delivery success/failure logs and the model now avoids certain postcodes on certain days purely from pattern recognition — things a human dispatcher would know intuitively but could never document formally. The RL agent uses a simple reward function: +10 for on-time delivery, -5 for late, -20 for failed attempt, and after about 60,000 episodes on historical data it started producing smarter routes than our senior dispatchers in blind tests. The paper that influenced our approach most is "Deep Reinforcement Learning for Vehicle Routing Problems" — searchable on arxiv.org — and it's worth reading even if you're not a researcher.

DanielTobin_Freight
Fleet Manager · Dublin
20 hours ago
One thing that burned us early on: we optimized purely for distance and our drivers were exhausted because the AI kept routing through areas with terrible parking and frequent stops requiring heavy lifting — metrics that weren't in the model at all. In 2026, if you're not including driver fatigue scores and load weight per stop as constraints, you're building a route planner that looks good on a spreadsheet and causes turnover on the ground. We now use a simple fatigue proxy: cumulative stop count × average parcel weight × time-of-day coefficient. It's not perfect science but it made a real difference in driver satisfaction scores within 3 months. Don't fall for the trap of optimizing only what's easy to measure.