Phase 1
Discovery
Two weeks with the pricing team and three weeks rebuilding their historical sales ledger into a clean training set. Most of the lift came from fixing the data, not the model.
A high-volume liquidation marketplace was pricing pallets from gut feel and a spreadsheet. We replaced both with a model that reads the photos, the manifest, and the prevailing market — and tells the buyer how confident it is.
A high-volume liquidation marketplace receives thousands of mixed pallets a week from national retailers — returns, overstock, shelf-pulls, and the occasional surprise. Each pallet has a manifest of varying honesty, a stack of photos taken in a poorly-lit dock bay, and a buyer somewhere who wants a price by tomorrow. Pricing analysts were doing it from memory and a colour-coded spreadsheet that nobody outside the team could read.
The result was the kind of variance that's invisible until you measure it. The same SKU mix could be priced differently by two analysts on the same shift; recoverable margin was leaking out the bottom of the distribution; and when a senior analyst left, six months of pricing instinct walked out with them.
Leadership didn't want to take pricing away from the analysts. They wanted a tool that priced the obvious pallets in the background and gave analysts time to think about the weird ones.
Photos in, manifest in, market signals in. Out comes a price, a confidence band, and the three comparable pallets the model leaned on. The analyst still owns the final number.
Reads the pallet photos and the manifest together, identifies the dominant categories, and estimates condition mix without being fooled by the shrink-wrap.
Maps each retailer's quirky SKU formats to a single internal catalog. Handles the cases where the manifest says "assorted" and the model has to do the work itself.
Retail price scrapers and resale-channel feeds piped into the model as features, with staleness tracked per source so the model knows when to widen its bands.
Conformal prediction intervals tuned on the operator's historical sell-through data. When the model says 80% confident, sell-through actually lands inside the band 80% of the time.
The pricing screen surfaces the model's number, the three nearest comparable pallets, and the features that mattered most. Analysts override with a one-line reason that becomes training data.
Same model, flipped: each registered buyer gets a daily list of pallets that match their historical purchasing profile, ranked by expected fit.
Two weeks with the pricing team and three weeks rebuilding their historical sales ledger into a clean training set. Most of the lift came from fixing the data, not the model.
One warehouse, half the inbound volume, four weeks. Analysts saw the model's price as a suggestion next to their own — and the override patterns gave us the second training pass.
Conformal calibration on a held-out quarter, SSO, role-aware pricing margins, and integration with the marketplace listing tool. The model's outputs are now signed and timestamped for audit.
Both coasts, the buyer-side recommendation feed live for the top hundred accounts, and a calibration dashboard the operator's analytics team now monitors without our involvement.
The marketplace's throughput ceiling used to be the pricing desk: pallets piled up on the floor waiting for an analyst to look at them. Six weeks after the full rollout, the pile was gone. Pallets cleared pricing in minutes, listed on the marketplace the same afternoon, and the buyer-side recommendations were doing real work — three of the top ten buyers were converting more than half of their feed.
The variance story was the one leadership cared about. The interquartile range of analyst-to-analyst pricing on a matched-pallet test set tightened by almost two-thirds, and the long tail of catastrophic underpricing — the pallets that sold in eight hours because the price was wrong — stopped showing up in the weekly review.
After the first warehouse, the operator brought us back for the buyer-side feed and a category-level forecasting workstream. Reference available — call us first.
Talk to a reference →"The thing that sold the analysts was the comparable pallets panel. They could see why the model said what it said. After two weeks nobody wanted to go back to the spreadsheet."
Model architecture, calibration methodology, and a redacted walkthrough of the analyst workbench — sent over after a short call.