Building a Semantic Model for Recycling Contamination Triage

Description and use

ResidueDesk‑AI uses a semantic model to evaluate loads at a recycling facility, considering both numeric metrics and prohibited words in notes, and produces a priority brief for the shift.

Problem

Recycling facilities need to decide how to handle inbound loads when a numeric residue percent is low but a note contains a prohibited item, such as a battery pack, which should override the numeric assessment.

Outcome

On a sample of ten tickets, the brief reported 2 accepts, 4 resorts, 3 diverts, and 1 reject, with an average residue of 8.36%. The system correctly quarantined a load with a battery note despite a low residue percent.

Workflow

• Validate ticket with Pydantic
• Project ticket to semantic layer (Load, Commodity, Supplier, measures)
• Scout for prohibited words in note
• Grade load against commodity metrics (ResiduePercent, MoisturePercent caps)
• Route load to appropriate line (accept, resort, divert, quarantine)
• Generate shift brief sorted by severity
• Expose API via FastAPI and CLI via Rich

Evidence

I built ResidueDesk-AI, a personal experiment, so a laptop at a fictional scale house would quarantine that load instead of waving it onto a tip floor because the residue percent was low.

Source

The PoC is a semantic model with a local coordinator on top. The model names three entities, Load, Commodity, and Supplier, and three measures on Load: NetWeightLb, ResiduePercent, and MoisturePercent.

Source

On the ten sample tickets the brief lands at 2 accept, 4 resort, 3 divert, 1 reject, average residue 8.36 percent.

Source