Inspiration
A single line in Trucker Path's own deck stopped us cold: "The product has the inputs. It doesn't yet have the layer that turns them into decisions." Maria manages 18 trucks using spreadsheets and phone calls. She spends 90 minutes every morning just finding out where her drivers are. At 11am a customer calls asking where their freight is — she doesn't know. At 5pm she chases one driver for a receipt to build one invoice by hand. Trucker Path already had all the data. Navigation, HOS clocks, tracking, document uploads — all of it. It just never triggered action. That gap is what we built LoadLorry.AI to close.
What it does
LoadLorry.AI is an autonomous AI agent that runs a fleet dispatcher's entire operational day across five challenges simultaneously:
Smart Dispatch — predictive 48-hour load assignment engine that scores drivers by HOS, location, fatigue, and return load probability. Maria's 90-minute morning becomes 90 seconds. 3D Digital Twin — live Three.js truck model where components pulse red when health thresholds are breached. Tire pressure drops below 80 PSI — the tire glows red before it blows out on I-10. Proactive Alerts — statistical process control filters 34 minor deviations as noise and surfaces only 2 decisions that actually matter. Every alert arrives with a pre-built action ready to approve. Billing & Document Automation — driver photographs BOL, POD, and fuel receipt. Claude Vision reads all three, reconciles against the rate confirmation, flags variances, and builds a complete invoice. 45 minutes becomes 10 seconds. Cost Intelligence — automated 3-way P&L reconciliation joins rate confirmation + telematics + signed POD the moment a trip closes. Plain English narrative tells Maria exactly why margins dropped and what to fix Monday morning.
Tagline: One agent. Five problems. Zero phone calls.
How we built it
Frontend: React + Vite with React Context for global fleet state. Five screens share one agent brain — approving a relay in the Alert screen simultaneously updates the Digital Twin, Dispatch Board, and sidebar badge. 3D Model: Three.js built from primitives — BoxGeometry for cab and trailer, CylinderGeometry for tires and exhaust stacks. Component colors are data-driven. Critical components pulse using: javascriptmesh.material.emissiveIntensity = 0.3 + Math.abs(Math.sin(Date.now() * 0.003)) * 0.6 Backend: Python FastAPI with endpoints for dispatch scoring, alert detection, OCR processing, and cost reconciliation. AI: Claude API (claude-opus-4-5) for Vision OCR — extracts structured data from driver document photos and builds complete invoice objects in under 10 seconds. Dispatch Engine: Custom predictive scoring algorithm: S(d,l)=0.30⋅H(d,l)+0.20⋅I(d)+0.20⋅T(d)+0.15⋅R(l)+0.15⋅F(d)S(d,l) = 0.30 \cdot H(d,l) + 0.20 \cdot I(d) + 0.20 \cdot T(d) + 0.15 \cdot R(l) + 0.15 \cdot F(d)S(d,l)=0.30⋅H(d,l)+0.20⋅I(d)+0.20⋅T(d)+0.15⋅R(l)+0.15⋅F(d) Platform: Trucker Path Fleet API for live driver, vehicle, and trip data with graceful mock fallback so demo never breaks.
Challenges we ran into
API Discovery — The Trucker Path platform serves an SPA that returns HTML for all routes. The real API lived on a different subdomain. We spent two hours in Chrome DevTools intercepting fetch calls to find it. Our solution: build a graceful fallback layer that tries every known endpoint pattern and falls back to realistic mock data. The LoadLorry.AI demo works either way. The 3D truck looked like a toy — Our first Three.js truck was colored boxes. We spent three hours on materials, lighting, and proportions. The breakthrough: realistic materials matter more than complex geometry. The right shininess, specular, and emissive values on simple shapes beat complex geometry with flat colors. Five features, one story — Five independent screens needed to feel like one agent. We designed shared React Context state first and built screens second. Every feature reads and writes to the same LoadLorry.AI fleet brain. 36 hours — We prioritized ruthlessly. The 3D twin never got cut. Everything else was built to support the 5-minute demo flow.
Accomplishments that we're proud of
The 3D truck tire pulsing red in real time when PSI drops below threshold — no other team built this The billing pipeline taking a photo to an approved invoice in under 10 seconds using Claude Vision Five completely connected screens sharing one agent state — approving one action cascades across the entire LoadLorry.AI dashboard A pitch that follows Maria's exact day from 7am to 5pm — judges felt the problem before they saw the solution
What we learned
The trucking industry runs on relationships and gut instinct accumulated over decades. AI doesn't replace that — it gives it a better interface. The difference between a tool and an agent: a tool answers questions. An agent acts before you ask. Alert fatigue kills adoption. Statistical process control that suppresses noise is more valuable than comprehensive monitoring. 17 hours of dispatcher time saved per week from billing automation alone:
23×(45 min−10 sec)≈17.2 hrs/week23 \times (45\text{ min} - 10\text{ sec}) \approx 17.2\text{ hrs/week}23×(45 min−10 sec)≈17.2 hrs/week
What's next for Load Lorry
Connect to live Trucker Path telematics API for real tire pressure and engine fault data Integrate with ELD providers (Geotab, Samsara) for real-time HOS feeds Add natural language interface — Maria types "who can take a Dallas load by 2pm?" and LoadLorry.AI returns an instant ranked answer Expand return load probability model using live TruckLoads board data Launch Margin Leakage Widget as free feature inside Trucker Path's existing fleet dashboard — converting 2,800 existing customers at $299/month
Log in or sign up for Devpost to join the conversation.