Problem
A logistics firm priced freight loads by hand. Could a model suggest quotes fast enough to help analysts win loads, without taking the pricing decision away from them?
Approach
A feedback loop, not an oracle: the model proposes a price from historical data; an analyst approves, makes a bounded correction, or rejects with a reason; that feedback retrains the model in batches.
What I built
- A pricing model that suggests a quote from historical load data.
- A review step where an analyst approves, makes a small bounded correction, or rejects with a written reason. Every decision is captured as labeled feedback.
- Batched retraining on the collected feedback.
- An evaluation of time-to-correct-price for analyst-alone vs. analyst + model.
Result
Measured how much faster an analyst reached the right price with the model in the loop. The hard parts were cleaning years of messy historical data and the minority load classes, which behaved unpredictably.