Astra is Shipsy’s autonomous planning agent. She takes the full pending-order book, the available fleet, the active SLAs, and the rolling real-world execution signals, and produces a daily plan — routes, sequences, vehicle allocations, slot commitments — that is 15–25% cheaper than what a legacy TMS planner produces, and published in minutes instead of hours. Astra does not replace the planner. She replaces the spreadsheet underneath the planner.
Why we built this
Planning is the most consequential decision in a logistics network, and it is usually made by people who are under-tooled. In every enterprise we’ve audited, planners run on three things: tribal knowledge of which driver handles which cluster, an Excel grid of capacities, and a stopwatch before the cut-off. The output is stable but not optimal — and it does not adapt when reality diverges from the plan.
Astra was built to turn daily planning from a spreadsheet-and-judgment exercise into a continuously optimized decision loop. She uses the combinatorial horsepower that human planners can’t, and she learns from how each day’s plan actually executed.
How it works
Astra runs three planning horizons, each with its own mechanism:
Horizon 1 — Day-of planning (micro-cluster routing). For each planning run, Astra builds micro-clusters from the pending order set — geographic clusters at 200–500m granularity, refined by service level, SLA, delivery window, and vehicle type fit. She then sequences within and across clusters using a variant of column-generation routing, tuned with 20 years of encoded courier tribal knowledge (parking-spot detection from accelerometer data, building-access time penalties, one-way-street avoidance, left-turn aversion in congested geographies). The result is route shapes that are demonstrably closer to what an experienced on-ground driver would draw than what a generic OR-Tools solver produces.
Horizon 2 — Allocation (driver-to-route, vehicle-to-trip). Astra matches routes to drivers and vehicles using four criteria: capacity fit (weight, volume, SKU compatibility), skill fit (B2B dock drivers vs residential drivers, cold-chain certification, white-glove training), territory familiarity (drivers perform 10–20% faster on familiar territory), and fatigue/duty-hour balancing (mandatory rest compliance, shift fairness). Allocations are re-evaluated per day, so no driver gets permanently stuck on an unfavorable territory.
Horizon 3 — Reactive replanning. Astra runs in the background during execution. When a trip goes off-plan — a reassignment, a driver break, a hub delay, a high-value exception — she triggers dynamic rerouting on the affected scope, publishes the updated task stack to the affected drivers, and recomputes ETAs for the consignee-facing tracking page.
Underneath all three horizons, Astra’s optimization model uses a multi-objective cost function — not a single “shortest total distance.” The weights balance: cost per stop, SLA breach penalty, driver overtime cost, customer CX impact (FADR, predictive ETA accuracy), and fleet utilization. The weights are tunable per enterprise and per lane; most enterprises run two or three profiles (cost-optimal, SLA-optimal, balanced) and let Astra pick per planning run based on the day’s risk profile.
Here’s the flow at a glance:
Early results
Enterprises deploying Astra typically report, within 60–90 days:
- 15–25% reduction in planning cost — measured as cost per stop or cost per shipment after allocation.
- Planning time collapses from 2–4 hours to under 15 minutes per run, so multiple replans per day become feasible.
- FADR uplifts of 3–7 percentage points because micro-cluster sequencing respects consignee availability windows better than distance-only routing.
- Overtime costs down 10–20% via duty-hour aware allocation.
A global big-and-bulky retailer leading in furniture and home goods uses Astra to plan dense metro delivery networks while maintaining 95% FADR on bulky scheduled slots.
What’s next
Three upgrades in flight: scenario planning (Astra produces three candidate plans per morning with explicit trade-offs so the planner can pick the risk profile), supplier-inbound coordination (plant-to-DC trips planned jointly with outbound last-mile so cross-dock windows align), and carbon-aware cost functions (routes optimized for emissions within a cost/SLA envelope).