Jev for Physical AI
Measured runs of Jev triaging incidents for a simulated 10,000-robot warehouse fleet, with cost and a ModernBERT comparison.
Overview
Jev for Physical AI by RoboKrunch applies Jev to robots, fleets and edge hardware with measured numbers from runs on 19 September 2026. Demo A asks whether Jev can be the decision layer for a warehouse AMR fleet: 300 incidents sampled from 41 bilingual Chinese and English templates (LiDAR degradation, localisation drift, battery faults, pallet misses, network partitions, human-zone intrusions), each answered with three judgements: escalate to a human, owning team, and urgency.
The README reports 300 of 300 successful calls via OpenRouter, 0.527 s median and 0.813 s p95 latency, about $24.57 per million decisions, and 91.3 percent agreement with template team labels, plus a monthly cost model for 14.4 million decisions compared with an estimated small LLM. Demo B compares Jev with a self-hosted ModernBERT to find the build-versus-buy crossover. Limitations are stated, including that templates are synthetic.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Should this incident be escalated to a human? | — |
| Choice | Which team owns this incident? | — |
| Score | How urgent is it (0 to 2)? | — |
How it works
- Sample incidents from bilingual templates
- Ask Jev three judgements per incident
- Measure latency, cost and agreement with labels
- Model fleet-scale cost and compare with self-hosted ModernBERT
Strengths
- Real latency and billing figures
- Bilingual incidents
- Build-versus-buy analysis
Caveats
- Synthetic incident templates
- Agreement measured against template labels, not real outcomes
Getting started
See the Reproduce section of the README.At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- Benchmark scripts
- Has code
- yes
Use cases and integrations
How curated lists describe it
Fleet triage: runs Jev as the decision layer for a 10,000-robot warehouse fleet over 41 bilingual incident templates and publishes 0.527 s p50 latency, $24.57 per million decisions and 91.3% agreement with template labels, alongside a crossover against a self-hosted ModernBERT.
yibie/awesome-jev
Frequently asked questions
What does Jev decide?
Whether to escalate an incident, which team owns it and how urgent it is.
What did it cost?
About $24.57 per million decisions in the reported run.
How accurate was team assignment?
91.3 percent agreement with template labels.
Related projects
Scores OpenTelemetry logs with Jev for diagnostic value, priority and routing before any expensive LLM analysis.
ProjectbetaTypeScriptBash-native CLI that filters any stream by meaning with Jev, like grep for semantics, so an agent wakes only for lines that matter.
ProjectbetaTypeScriptOpenTelemetry Collector processor that asks Jev how useful each metric is from its metadata and applies keep/drop policy.
ServiceprototypeGoOpenTelemetry trace processor that asks Jev whether an operation is diagnostic, business-critical and worth keeping, and annotates spans.
ProjectprototypeGoROS 2 robot diagnostics CLI where an optional agent uses Jev to choose which read-only evidence tool to call next.
CLIbetaPythonSimulated Franka arm that follows plain-English goals by having Jev chain about 30 hardcoded motion primitives, one per step.
Agent skillprototypePython