Haanji!

← All agents

🎙 Test in browser →Flow editor →

Dialed

Connected

100%

Qualified

Meetings

Avg duration

1m 2s

Total talk time

🧪 Prompt A/B test· primary config vs alt_config

Variant A

primary

Calls5

Qualified rate0%

Avg judge score4.3/10

Variant B

candidate

No calls routed yet.

⚙ A/B test setup· no A/B configured

Create a candidate prompt variant and route a slice of calls to it. The dashboard tracks both variants' qualified rate and judge score so you can pick a winner statistically — see the A/B test card above for live results.

Product

Agricultural Machines & Tools

बलवान कृषि

Languages

hi-IN

Qual fields

2 guardrails

Max call

5min

Recent calls

All calls →

When	Campaign	To	Duration	Score	Stage	Disposition
7/25/2026, 9:28:04 AM	—	+919267935682	40s	0	connected	Not qualified
7/14/2026, 7:18:28 AM	—	+919782000788	53s	0	connected	Callback (qualified)
7/13/2026, 8:41:41 AM	—	+919602582627	43s	0	connected	Not qualified
7/2/2026, 11:26:12 AM	—	+918560092222	1m 3s	0	connected	Not qualified
6/16/2026, 1:59:29 PM	—	+919461594523	1m 49s

Speculative LLM (latency boost)

Race a fast draft model against the primary. ~1s faster TTFA on slow models. Shadow mode collects metrics without changing caller audio.

Enable speculative draft+verifyShadow mode — collect metrics only, do NOT emit draft to caller

Draft providerDraft model (optional)

Match threshold: 0.85Cosine ≥ this value keeps the draft. Lower = more aggressive (more latency wins, more pivots).Hard pivot on mismatch (interrupt TTS mid-word — accurate but can sound jarring)Skip speculation in OBJECTION state (recommended — preserves precision)

Quality features (per agent)

Override the org-wide env flags for this agent. "Default" inherits from env; "On" / "Off" force the behaviour regardless of the env value.

Multi-agent debate

Closer-vs-skeptic parallel personas fire on sticky objection / low confidence / OBJECTION state. ~1.5s extra LLM budget per triggered turn.

Self-reflection (inner monologue)

Per-turn hidden JSON thought fed into the next turn. Runs async during TTS playback so zero added latency, but costs 1 extra LLM call per turn.

LLM-as-judge post-call scoring

After each call, score the transcript on the active rubric and write to call_scores. One extra LLM call per completed call.

Voice biometric fraud check

Embeds caller voice + scores against org blacklist + deepfake detector at call start. Skip on low-value retention agents to save ~80ms per call.

Advanced sub-knobs (override env defaults)

Debate low-confidence threshold (0–1)Debate hard budget (ms)Reflection every N turns (≥1)LLM hard timeout (sec)

Whisper STT (only if STT provider = whisper)

⚠ Bot loads ONE model per process — first agent to dial wins. Restart bot to swap.

ModelDeviceCompute type

Call quality & turn-taking (advanced — blank = bot default)

How the bot listens and how patient it is with silence. Lower VAD stop/start = snappier but more likely to cut the caller off; higher = more patient. Idle timeout + reprompt count control how long it waits on a quiet line before giving up.

Idle reprompt timeout (sec)dead air before the bot re-checks · default 15Silence reprompts before hangupdefault 3 · higher = more patient with a quiet callerVAD stop (sec)silence before a turn ends · default 0.45VAD start (sec)speech needed to trigger · default 0.3VAD confidence (0–1)speech-detection strictness · default 0.7VAD min volume (0–1)background-noise floor · default 0.6Echo gate cooldown (sec)mute mic after bot speaks · default 0.4

Knowledge base

The product facts this agent may state on a call. Each entry is embedded and surfaced to the agent by the RAG retriever when it's relevant.

Loading…

Title

Smart specialists

Subagent specialists handle specific moments of the call alongside the main agent. Each is opt-in. The main agent always has a working fallback — if a specialist times out, the call continues normally.

Objection handler

When the bot detects a prospect objection (price, timing, trust, fit), a specialist crafts a tailored response using your objection bank + recent transcript. Soft fallback to classic logic on timeout — zero risk to live calls.

Qualifier (Wave 3)

A parallel watcher that runs every 3rd turn — predicts qualification score, flags missing fields, suggests next question. Async / soft hint; never blocks the bot.

Closer (Wave 4)

When the prospect signals buying intent OR just crossed the qualification threshold, a specialist crafts a single-turn close pitch with a specific next-step ask. Fires once per call. Soft fallback to classic logic on timeout.

All fields look good

Cancel