Insights
AI traffic audit results for your workspace.
Xybrid maps your AI traffic into workflows, then finds what is expensive, unsafe, or broken.
Workflows discovered
Workflows explain what keeps happening.
Prioritized findings
Findings explain what should change.
- critical Security Possible secret sent to AI provider
An API-key-shaped string was detected in 3 prompts forwarded to OpenAI in the last 24 hours.
→ Redact credentials before they reach the provider, or route the workflow through the gateway’s redaction filter.
- high Cost Expensive model used for a short-output workflow ~$42.80/mo
gpt-4o handles ~8.4k calls/day that average 40 output tokens; a smaller model matches quality on this shape.
→ Switch support-reply-generation to gpt-4o-mini — projected ~$42.80/mo saved at current volume.
- high Quality Refusal rate climbing on a production workflow
12% of moderation calls returned a refusal in the last 7 days, up from 3% the week prior.
→ Review the system prompt — refusals spiked right after the prompt change on the 3rd.
- medium Cost Duplicate prompts not being cached ~$11.20/mo
26% of faq-bot prompts are exact duplicates within a 5-minute window.
→ Enable prompt caching on faq-bot to cut the redundant spend.
- medium Latency Workflow has high p95 latency
p95 is 8.9s against a 2.1s median — long-context calls dominate the tail.
→ Cap context to the last 20 messages, or split the summarize step.
- low Routing Cloud fallback firing more than expected
9% of calls fell back to cloud last week against a 2% target.
→ Investigate on-device failures for the whisper stage.
Evidence and triage are available once you sign in.