Blog

Jev scores clean calls well, but it can't abstain. See where Jev accuracy breaks on real calls, how forced answers skew QA metrics, and how to fix it.

Break a 30-minute manual test call session into 16 checks, then automate them in four layers on LiveKit or Pipecat, with cost math and flake control.

Voice agent prompt versioning for LiveKit and Pipecat: pin each call to a release, gate on pass^k, size the canary, and roll back without dropping calls.

Flux or Nova-3 for a phone voice agent? Verified features, pricing at 50k to 500k calls, what each latency number means, and a bake-off you can run.

Record LiveKit and Pipecat calls in stereo: Egress channel routing, AudioBufferProcessor, Twilio dual-channel, clock alignment, storage math, metric code.

LiveKit agent Kubernetes guide: how agent servers decide they're full, cgroup load traps, drain math for live calls, job-based autoscaling, safe canaries.

What runs on self-hosted LiveKit for noise and echo: Krisp, BVC, open-source filters, AEC warmup, and a test plan with SNR beds, WER and barge-in rates.

Measure Pipecat latency stage by stage: what each TTFB times, how UserBotLatencyObserver splits a turn, and how to check it against a call recording.

Silence timers, Deepgram Flux, LiveKit and Smart Turn models, and semantic VAD compared: defaults, published benchmarks, the stacking trap, and a test.