Relayiq V2 — Productivity skill for Claude Code

RelayIQ v2 — real-time voice agent on Twilio, Deepgram, Claude, ElevenLabs

A telecom customer-care voice agent (billing, account status, outages, plan questions) on a streaming architecture, answering a live PSTN call with a caller-to-first-syllable latency of **2.9 s median (2.1–6.3 s)** across a 3-turn live call, with the **25-case eval suite passing its release gate on Claude Haiku 4.5** (agent p95 3.9 s) and every input/output guardrail decision costing **under 30 ms**.

Result

Live call — Twilio → Deepgram → guardrails → Claude → guardrails → ElevenLabs

First real call through the full stack, September 22, 2026 (`docs/evidence/first-live-call-2026-09-22.log`):

Turn	guardrail_in	agent	guardrail_out	TTS first byte	First byte to caller	Within 1.2 s budget
1	28 ms	1,888 ms	18 ms	+170 ms	2,104 ms	no
2	8 ms	2,706 ms	21 ms	+212 ms	2,948 ms	no
3	8 ms	5,606 ms	12 ms	+717 ms	6,343 ms	no

The agent stage is 90–95 % of every turn. Guardrails and TTS are not the problem; the LLM turn is. The output guardrail rewrote one reply before it was spoken (`action: modify` — voice-format layer). Three of three turns resolved without escalation.

Eval suite — 25 labeled cases through the production pipeline

Run	Model	Deterministic checks	Agent p95	Gate
1	Claude Sonnet 4.5	20 / 25	—	fail — 4 wrong test expectations, 1 real grounding bug (ISO datetime regex)
2	Claude Sonnet 4.5	25 / 25	5,933 ms	fail — latency threshold (≤ 4,000 ms)
3	Claude Haiku 4.5	24 / 25	3,931 ms	pass — one over-escalation: grounding blocked a model-computed total

Run 3 is the shipped configuration (`ANTHROPIC_MODEL=claude-haiku-4-5`). Full per-case output: `eval/results/`. Thresholds: `eval/thresholds.yaml`.

Architecture — the data path

caller's voice ──► PSTN ──► Twilio number