Atlas, Shipsy’s autonomous control tower, is designed to resolve incidents before a human sees them. Every shipment and vehicle is watched by detection rules that fire on state + signal + prediction; detected incidents are routed to the right owner (driver, hub lead, carrier, merchant, or an agent), and 60-80% are auto-remediated by one of Shipsy’s four agents — Clara, Nexa, Vera, or Astra — before the human queue ever gets longer.
Why we built this
Control towers have existed for 15 years, and most are passive dashboards. Someone stares at a screen, reacts when a red dot appears, and shouts into the radio. That model collapses at enterprise scale — a global parcel leader spanning 65+ countries runs 100K+ trips a day, and no dashboard-watching team scales to that. A global pharma CDMO running multi-country clinical supply has even tighter latency constraints because a 4-hour detection delay can destroy a shipment.
Atlas replaces dashboards with outcomes. It detects, it decides, it acts. Humans are pulled in only for genuine judgment calls.
How it works
Detection layer. Atlas ingests a multi-stream event feed: GPS pings, milestone events, driver-app telemetry, IoT sensor data (temperature, door-open, tilt), carrier webhooks, customer signals (NDR, reschedule requests), and internal state changes. Detection rules combine three ingredients — current state, live signal, and predicted state. Examples:
- Geofence-miss: vehicle has been within 500m of a consignee for 20+ minutes but no delivery event has fired → potential fake-delivery or blocked access.
- Predictive ETA breach: Astra’s predicted ETA has slipped past SLA by more than 15 minutes → SLA breach is imminent even though it hasn’t happened yet.
- Cold-chain excursion: temperature outside acceptable band for 8+ continuous minutes → remediation needed before product damage.
- Silence anomaly: no GPS ping for 45 minutes on an active shipment → driver health, device failure, or deliberate diversion.
Predictive detection is the unlock. Most legacy control towers detect after the SLA is missed. Atlas detects the drift that will miss it, giving the remediation path 30-90 minutes of lead time.
Routing layer. Each incident type has a routing table: who owns it, which agent tries first, what the escalation ladder is. A predictive-ETA breach on a consumer parcel routes first to Clara (who proactively messages the consignee offering slot change) and Astra in parallel (who re-sequences the route if a faster save exists). A cold-chain excursion on a pharma shipment routes instantly to the hub lead and the 3PL carrier contact — there’s no agent auto-remediation for a pharma incursion because compliance demands a named human sign-off.
Auto-remediation via agents. Where policy allows, the incident is handed to the appropriate agent:
- Clara handles CX-facing incidents: NDR rescue (re-attempt slot negotiation with consignee), proactive delay comms, refund or reattempt decisions within a bounded refund threshold.
- Astra handles operational incidents: dynamic re-routing, driver swap, carrier swap, depot re-sequencing. See dynamic rerouting during execution.
- Nexa handles financial/billing incidents: invoice mismatches, failed deposit matching, rate-card misapplication.
- Vera handles dispute-driven incidents: carrier settlement disputes, damage claims, merchant payout disagreements.
Escalation discipline. Every incident carries an SLA-to-resolve clock. If an agent can’t close the incident within a configured window (usually because the situation is outside the agent’s bounded authority), the incident is escalated to a named human queue with the full context pack: event trail, predicted outcome, options explored, recommended action. Humans see one-click approval screens, not investigation tickets.
Situation synthesis. When multiple correlated incidents fire in a short window (e.g., a hub-scan failure plus a batch of predictive-ETA breaches for shipments from that hub), Atlas rolls them into a single “situation” with the root cause flagged. This prevents the human team from drowning in 40 tickets that are really one problem.
Audit trail. Every detection, routing decision, agent action, and human approval is event-logged. For regulated shippers (pharma, defense, high-value goods), this is the audit record that replaces the ad-hoc ticket system. A global biotech scaling rare-disease therapeutics uses this trail for regulatory traceability across 30+ countries.
Early results
Enterprises running Atlas report: 60-80% of incidents closed without human involvement (for CEP and quick-commerce), SLA recovery rates doubling because predictive detection gives the save-path real lead time, and “control tower headcount” redeployed from dashboard-watching to exception-judgment work. A major beverage bottling group running primary + secondary distribution moved from a 12-person control room to a 4-person exception desk after Atlas’s first year, while throughput grew. See a related analysis on control tower evolution.
What’s next
Next release introduces situation-level learning: Atlas remembers which remediation paths worked for a given incident signature and biases toward them on next occurrence, raising auto-remediation share over time. Human approvals feed the same learning loop.