Question 1

How do you build an HOS-aware driver-pay and load-acceptance model without violating ELD rules?

Accepted Answer

The model is constrained, not just predictive. Inputs come straight off the ELD: remaining 11-hour driving time, 14-hour on-duty window, 30-minute break status, 60/70-hour 7/8-day cycle, sleeper-berth splits, plus the load profile (pickup/delivery appointments, deadhead miles, lane RPM, accessorials). A mixed-integer programme maximises driver-pay-per-week subject to FMCSA 49 CFR 395 limits as hard constraints — never recommended slack. We also model the CSA scoring impact of marginal violations and treat any score-tipping load as infeasible. Driver acceptance probability is learned from historical tender behaviour by driver cohort. The output for dispatch is a ranked tender list with the HOS feasibility window, expected detention, and a margin estimate.

Question 2

How do you ingest X12 204 / 214 / 990 / 856 into Manhattan Active TMS without breaking the nightly cycle?

Accepted Answer

The ingest path is parallel to the TMS, not in front of it. EDI lands on an AS2 or VAN connector, gets parsed (Smooks / EDIFACT-X12) into a canonical shipment event with full envelope retention for the 7-year retention regulators expect. We push the canonical event into a Kafka topic and into Manhattan Active TMS via its REST endpoints during a documented write window outside the nightly rate engine, planner and EOD jobs. The 214 status events flow into the visibility layer in real time so customer-facing ETA does not wait on TMS batch. 990 acceptance/rejection goes back to the broker in under 30 seconds — most brokers contractually require under 5 minutes. For Blue Yonder Luminate and Oracle OTM the pattern is the same; only the adapter changes.

Question 3

How accurate is your ETA prediction when you fuse FourKites + weather + carrier-historical signals?

Accepted Answer

On regional truckload lanes under 500 miles we hit sub-1 hour MAE measured against the actual gate-in timestamp at the receiver. On transcontinental TL it is 2–4 hours depending on lane volume. Feature set: FourKites/Project44 GPS pings (cleaned for ping-density gaps and spoofing), origin/destination dwell distribution by facility (we maintain a facility profile per DUNS), HRRR/RAP weather along the planned route, HERE traffic forecast at the predicted arrival window, carrier-historical OTP on that lane, day-of-week and shipper appointment-honour rate. The model is a temporal fusion transformer for the trunk, a quantile regression head for the arrival window, and a calibration layer (isotonic) so the stated 80% confidence interval actually contains the actual arrival 80% of the time — most carrier ETAs are not calibrated and that is what breaks downstream appointment planning.

Question 4

How do you detect double-brokering and identity fraud at carrier onboarding?

Accepted Answer

Double-brokering is a graph problem disguised as a KYC problem. We pull FMCSA SAFER (DOT number, MC number, authority status, BOC-3, insurance filings), CSA scoring, the carrier’s declared address, banking details, phone, email and W-9, and we resolve all of those against a carrier graph built from prior onboarding attempts, paid-invoice history and known-fraud lists (CarrierWatch, Highway, Truckstop.com Carrier Monitor, plus internal blacklist). Signals that flag: MC number flipped to a new entity within 90 days, address shared with a recently-deactivated carrier, banking ACH details matching a previously-charged-back account, phone or email reused across two MC numbers, insurance certificate issued by a non-existent producer. A gradient-boosted classifier on those features plus a graph neural network on the relationship graph; we run it pre-tender, not post-load-board-match, so the fraud never gets a load assigned in the first place.

Question 5

How do you calculate scope-3 emissions per shipment using GLEC and ISO 14083?

Accepted Answer

The GLEC Framework v3.1 mapped onto ISO 14083:2023 vocabulary. For each shipment leg we compute energy intensity (g CO2e / tonne-km) using the carrier-specific factor where we have it (carriers in SmartWay, EcoTransIT subscribers, or where we can derive intensity from ELD fuel-burn + payload), otherwise the GLEC default by mode-fuel-region. We respect well-to-wheel (WTW) versus tank-to-wheel (TTW) distinction explicitly because CDP and SBTi reporting require WTW. Allocation is mass-based for full TL/FTL and by chargeable-weight for LTL and parcel. Output is per-shipment CO2e, CH4 and N2O, attributable miles, the chosen emission factor with provenance, and a confidence band. The result feeds CDP disclosure, SBTi targets, and customer carbon dashboards — we never invent factors and every number has a citable source for assurance.

Question 6

How do you optimise drayage and intermodal lane planning under port congestion?

Accepted Answer

Drayage is constrained by three things in this order: chassis availability, terminal appointment, and driver HOS. Optimisation is a rolling-horizon MIP. We ingest terminal appointment data (PierPASS, eModal, Voyager Track at LA/LB; APM Maersk Mobile and similar at NY/NJ), chassis pool status (TRAC, FlexiVan, DCLI), street-turn opportunities, and a dwell prediction at the terminal (which uses the terminal’s queue length and the historical appointment-to-gate-out distribution). The model schedules pulls, street-turns and per-diem returns to minimise total cost while keeping drivers inside HOS and respecting demurrage and detention clocks. We also model the BCO’s free-time clock explicitly because most cost overruns come from missed free-time, not from inefficient routes.

Question 7

How accurate is document AI on photographed BOLs, PODs, customs forms and carrier invoices?

Accepted Answer

On clean photographed BOLs we extract shipper, consignee, BOL number, PRO, weight, piece count, freight charge, hazmat indicator, and signature presence at over 97% field-level accuracy. On wrinkled, low-light driver-phone photos of PODs the accuracy on the four high-value fields (POD timestamp, signature presence, exception/OS&D code, consignee name) is 92–94%. The stack is a layout-aware transformer (LayoutLMv3-class) fine-tuned on the customer’s actual document corpus — generic OCR plus regex fails on BOLs because every shipper formats them differently. We always emit a per-field confidence and route low-confidence fields to a human review queue rather than auto-posting to the TMS, which is what AP/AR teams need for audit defensibility on freight invoices and detention/accessorial claims.

Question 8

How do you predict dwell time and detention at shipper and receiver facilities?

Accepted Answer

A per-facility dwell model keyed on DUNS / address-normalised facility ID. Features: day-of-week, hour-of-arrival, appointment type (live unload vs drop), commodity, equipment, season, the facility’s historical dwell distribution, the queue depth derived from FourKites/Project44 truck density at the geofence, and weather. We fit a survival model (accelerated failure time) rather than a regression because dwell is heavy-tailed and you care about P50, P90 and P99 differently. Output is plugged into appointment planning (tender to a different carrier if expected dwell exceeds detention-trigger threshold for the contracted driver), driver-pay (auto-detect detention-eligible time without waiting for the carrier invoice cycle), and the ETA model upstream so downstream appointments do not cascade. On RPM (rate-per-mile) lane forecasting for spot pricing we use the same facility-aware dwell as a feature alongside DAT and Truckstop posted volumes, lane velocity, and macro fuel/CPI indices — published RPM forecasts have 7–10% MAPE on weekly aggregate at the major-lane level.

Carrier-grade AI: sub-1h ETA MAE, X12 EDI into Manhattan Active, scope-3 per shipment

Platforms we integrate to

Six AI systems we build for 3PLs, brokers & carriers

ETA prediction & arrival-window calibration

Double-brokering & carrier-onboarding fraud detection

Document AI on BOLs, PODs, customs forms & freight invoices

Scope-3 emissions per shipment (GLEC + ISO 14083)

HOS-aware driver-pay, load-tendering & dwell prediction

Drayage, intermodal & last-mile optimisation

Carrier-grade data infrastructure

EDI fluency: X12, EDIFACT, GS1 EPCIS

Real-time visibility fusion

TMS / WMS CDC & event streams

ELD telematics & time-series at scale

Scope-3 emissions data fabric

Deployment topology for driver PII & customer data

What 3PLs, brokers & carriers actually ask us