Sigmalion
All case studies
ConfidentialLogistics · Poland · 400 employees

AI-powered route optimization rescued from 8-month delay

A Polish logistics company's in-house AI project had stalled for 8 months with nothing to show. The lead engineer left, the codebase was untested, and the board was ready to pull the plug. We took over, rebuilt the core engine, and shipped in 6 weeks.

12×
faster processing
−38%
fuel costs in pilot
6 wks
to delivery
01

Challenge

What was broken

A 400-person Polish logistics company had been running an internal AI project for route optimization for 8 months. The goal was to reduce fuel costs and dispatcher workload using ML. The lead engineer had just resigned, the two remaining developers were junior and struggling, and the model performed no better than their existing manual process. The board had given a 6-week ultimatum: ship something real or cancel the project and write off the investment.

02

Our Approach

How we thought about it

We started with a 48-hour codebase and data audit. The core problem was immediately clear: the ML model was heavily overfit to synthetic training data and failed on real-world route edge cases — irregular pickup windows, driver availability constraints, fuel stop requirements. The previous approach was trying to solve a constraint satisfaction problem with a pure ML model, which was the wrong tool. We proposed a hybrid: LLM-powered constraint parsing with classical optimization algorithms (modified Clarke-Wright) for the actual routing.

03

Solution

What we built

We rebuilt the optimization engine from scratch using LangChain for natural-language constraint parsing (dispatchers could now describe constraints in plain Polish) and a modified vehicle routing algorithm for the actual scheduling. FastAPI served the new backend, integrated with their existing PostgreSQL fleet database. We extended their React dispatcher dashboard with a new route visualization panel. The whole system was containerized and deployed to their existing on-prem servers to avoid any infrastructure change approval delays.

04

Results

What shipped

The system went live at the end of week 6 with a 2-week pilot on 12 of their 60 routes. Processing time dropped from 4 minutes per route to under 20 seconds. Fuel costs on the pilot routes fell 38% in the first month. The board approved full rollout and added budget for a second phase extending the system to cross-border EU shipments.

Architecture

System overview

Dispatcher Input
Natural language constraints
LangChain Parser
Constraint extraction
Route Optimizer
Clarke-Wright algorithm
FastAPI Backend
PostgreSQL integration
React Dashboard
Real-time visualization
PythonFastAPILangChainPostgreSQLReact
We had written off the project as a total loss. Sigmalion walked in, understood the problem within days, and shipped something we're genuinely proud of. The board went from cancellation to expansion budget.
CTO — Logistics SaaS, Poland (confidential)

Got a similar challenge?

Let's talk about your situation — 30 minutes, no commitment, and you'll leave with a clearer picture of how to move forward.