Welcome to Soulax

Menu

Stop guessing. Start claiming.​

For refrigerated carriers, dwell time, detention fees, and cold-chain risks are hidden in disconnected systems—until now. Soulax’s Trailer Analytics AI unifies your dispatch, telematics, and tractor data into a single, conversational platform. Turn messy data into actionable answers in seconds, not days.

At a glance

Client

N. American refrigerated carrier

Industry

Cold-chain freight / logistics

Fleet in scope

~650 reefer trailers · 450 tractors

Operation

~70% drop-and-hook

Platform

Amazon Bedrock

Pattern

Multi-agent orchestration

Data sources

AS/400 · ORBCOMM · GeoTab

Interface

Natural-language chat

The challenge

Trailers spend most of their life outside the carrier's control
In drop-and-hook operations, a refrigerated trailer is dropped at a shipper or receiver and picked up hours — sometimes days — later. For a carrier running roughly 650 reefer trailers with about 70% of moves drop-and-hook, that means most of the fleet is sitting somewhere the carrier can’t directly see, accruing dwell, detention exposure and cold-chain risk. The signals that explain what actually happened are real, but they are scattered across three systems that rarely agree:
  • Dispatch (ICC AS/400) — orders, appointments and load associations, often keyed by hand.
  • Trailer telematics (ORBCOMM) — reefer state, temperature and trailer location, streaming continuously.
  • Tractor & driver activity (GeoTab) — power-unit movement and hours, on a different clock again.
The result
Questions a dispatcher or account manager needs answered in the moment — “how much detention can we claim from this shipper?”, “why is this facility always slow?” — became data-team tickets that returned days later, when the load was long gone. The answers existed. They were simply unreachable at operational speed.

The Solution

Trailer Analytics AI — a fleet teammate you can talk to

Soulax built Trailer Analytics AI: an AWS-native platform that lets carrier teams ask fleet questions in plain language and receive an answered, reasoned, evidence-backed response — with inline charts, maps and journey timelines. Behind the conversational surface, the system unifies the three data sources into one time-reconciled model of the fleet and applies the carrier’s own business rules to every answer.

Crucially, it is not a single large language model responding to prompts. A detention claim or a compliance flag is too consequential for a best-guess answer. So the platform is built as a coordinated set of agents that plan, retrieve, reason and verify — the agentic pattern described below.

How it works · Agentic AI pattern

Five governed stages, not one black box

Every request is decomposed into a pipeline. An orchestrator decides intent, specialist agents own each domain, a knowledge graph holds the business rules, and a confidence gate decides when a judgment call warrants the foundation model versus deterministic logic. Each stage is traceable.

01

Orchestrator agent

PLAN & ROUTE

Interprets the request, resolves entities (trailer, account, facility, date range), classifies intent, and dispatches to the specialist agents needed. Multi-step questions become an ordered plan rather than a single guess.

02

Specialist agents

DOMAIN REASONING

Each agent owns one part of the fleet's reality and knows how to compute it correctly, including the edge cases the raw data hides.


dwell & benchmark detention & billing utilization & empty-move cold-chain compliance facility intelligence ETA & journey
03

Knowledge graph — semantic layer

GOVERNED RULES

Business logic isn't buried in code. Detention thresholds, dwell benchmarks and compliance rules live as versioned, auditable nodes — so an answer can be traced to the rule that produced it, and rules can change without a redeploy. Identity is modeled as type on the node, role on the edge, which is what lets one shipment be stitched across trailer, tractor and order.

04

Event detection & confidence gate

STREAM + GEN AI

A streaming layer detects load, unload and phase transitions from live telematics and resolves the clean, deterministic cases at scale. Ambiguous events — conflicting timestamps, unplanned stops, reefer-off, setpoint mismatches — are escalated to a foundation-model classifier for a judgment and a confidence score. Deterministic where it can be, generative only where it must be — which keeps cost and latency down without sacrificing accuracy.

05

Grounded response

ANSWER + EVIDENCE

The orchestrator composes the specialists' results into a single answer — prose, inline charts, maps and timelines — always tied back to the source events and the rule that was applied. Every insight is auditable, not a black box.

WORKED EXAMPLE

"How much can we claim from that shipper for detention last month, and where?"

The orchestrator routed the request to the detention and facility agents, pulled arrival and departure events from trailer telematics, applied the contracted free-time rule from the knowledge graph, reconciled it against dispatch records, and returned an itemized, evidence-backed claim across all facilities — with the worst-offending distribution centers ranked and the supporting timestamps attached. An answer that used to take a data team days arrived in the conversation.