Phase 1
Discovery
Two weeks on the plant floor with the sort team. We labelled a starter dataset on three streams and pinned down which contamination categories actually drove rejections.
A regional materials recovery operator was losing margin on rejected loads. We put cameras and a grading model on the conveyor — and put the sorting team back in front of the screens that actually matter.
A regional manufacturer of recycled materials runs a high-throughput MRF and ships baled streams to converters across the country. Contamination grading — the call on how much non-target material sits in a bale of HDPE or OCC — was being made by line sorters under fluorescent lights, on a conveyor moving faster than anyone wanted to admit. The accuracy was fine on average and terrible on the tails.
The tails were what hurt. A rejected load came back at the operator's cost, plus reload, plus the buyer's patience. Three of their five biggest accounts had started running their own incoming-load QA, and the rejection rate on a few SKUs had crept past what the commercial team could absorb.
They had looked at off-the-shelf sorting systems. Two problems: the systems graded by composition but didn't track grade against supplier, and the alerting was a flashing light at the end of the line — too late to do anything about it.
A vision pipeline that grades each stream in real time, routes contaminants where the existing sorters can catch them, and writes a per-supplier scorecard the commercial team can argue from.
Industrial line-scan and area cameras mounted over each conveyor segment, with stable lighting and a calibration target the maintenance team can re-run in under a minute.
Jetson devices running the grading model on-prem. No round-trip to the cloud, no dependency on the plant's internet during a shift.
A segmentation model trained on the operator's own material mix, retrained on a monthly cadence as feedstock composition shifts with the seasons.
A floor-mounted display per line showing current grade, drift against target, and which infeed batch is responsible. Built to be read at fifteen feet.
Every inbound load is graded and attributed back to the hauler and the source contract. The commercial team gets a weekly export they can actually negotiate with.
Real-time triggers to the line lead's radio when a stream drifts out of spec, with a clip of the offending frames attached for the post-shift review.
Two weeks on the plant floor with the sort team. We labelled a starter dataset on three streams and pinned down which contamination categories actually drove rejections.
One conveyor line, one shift. The model ran in shadow mode for three weeks, with daily reviews against the sort team's calls. By the end, agreement was tighter than human-to-human inter-rater.
Edge deployment, MQTT into the plant's existing OT bus, automatic failover to a known-good model if a camera fault is detected. A maintenance runbook the plant team actually owns.
Four additional lines, the supplier scorecard for the commercial team, and a retraining pipeline the plant's data lead now operates with a weekly check-in from us.
Within a quarter, the rejection log on the operator's three largest accounts had gone from a weekly conversation to a monthly one. The grading didn't change — the visibility did. When a load was off-spec, the team knew which infeed batch had caused it, which supplier had delivered the infeed, and what the contamination category was, all before the bale left the floor.
The commercial team's first real win came eight weeks in: a hauler whose loads were consistently grading 3 points below contract was renegotiated, with the scorecard as the evidence. That single conversation paid for the project. The sort team, meanwhile, stopped being graders and became exception handlers — a job most of them found materially less awful.
After the first plant landed, the operator brought us back for a second site and a supplier-onboarding workstream. The reference is available — call us.
Talk to a reference →"Most vendors hand you a dashboard and a phone number. East Reach handed us a model, a retraining pipeline, and a sort lead who knew how to run both. That's the difference."
Hardware bill of materials, model architecture notes, and a redacted week of the supplier scorecard — sent over after a short call.