RTS Labs built a proof of concept for a high-throughput recycler that needs to catch unsafe behavior on its conveyor lines the moment it happens. The result is a real-time computer vision system that watches every station continuously and dispatches AI agents to document and route each event into Field1st, RTS Labs' own field-safety platform.
High-Throughput Recycler (Proof of Concept)
Vision AI & Agentic Safety Automation
NVIDIA DeepStream
RTMPose + Triton
ByteTrack + OpenCV
Field1st
Our engineers will map your workflow and define a ship date in a 2-week Discovery Sprint.
On a material recovery sort line, safety is hard to see and harder to prove. Workers move fast alongside conveyors, balers, and heavy machinery, and the riskiest behaviors, reaching across a running belt, crowding a neighbor, stepping into an exclusion zone, happen in seconds. Sort-line safety programs typically lean on supervisors hitting an observation quota across a set of routes, so most of the line and most of every shift goes unwatched. Manual inspections take 30 to 35 minutes each and depend on memory, familiarity, and judgment, which means calls are slow, inconsistent, and easy to miss.
Existing telematics cameras push some alerts but catch almost none of the near-misses on the fixed sort lines. And nobody can measure the exposure that actually drives most injuries in waste handling: repetitive over-reach and torso lean, sustained across a shift.
We built the proof of concept around one priority: detect and warn in real time, while the risk is still happening. The architecture is also designed to run fully on-site, so a facility that needs it can keep detection and alerts working even without an internet connection.
Coverage depended on supervisors meeting an observation quota across 10 to 15 routes, so most of the line and most of every shift went unwatched.
Manual safety inspections took 30 to 35 minutes and leaned on memory, familiarity, and bias, producing inconsistent calls that were easy to miss.
Strains and sprains are the most common waste-handling injury, driven by repetitive over-reach (~80cm to the belt midpoint vs a ~50cm safe limit), yet nobody could quantify it across a shift.
RTS Labs built a three-layer system that keeps the intelligence at the edge and only reaches for a large model when it matters. Neural pose models read the scene in real time on an on-site device; a deterministic geometry-and-rules engine grades what they see, so every call is explainable rather than a black box; and when a violation fires, AI agents take over to confirm, document, and route the event into Field1st. Because the vision events share Field1st’s existing safety taxonomy, each detection lands as a record the team already knows how to action and feeds the next day’s job briefing.
Stop strategizing. Start building. Let’s map your workflow and get your AI integration into production in 90 days.
