Quick start
| I want to… | Go here | Password needed? |
|---|---|---|
| See a zero-cost scripted demo of a full call | /demo | No |
| Run a live call as a human agent (voice twin + coach) | /app | Yes — on "Start call" |
| Talk to a live AI agent as the customer | /agent | Yes — on "Start agent call" |
| Run a two-party live call (agent on one device, customer on another) | /app (agent) + /call (customer) | Yes — both devices |
| See the manager view: leaderboard, compliance %, human-vs-AI | /manager-preview | No |
| Learn what the product is | / and /about | No |
Access password (all live-call surfaces): provided on request — reach out via LinkedIn.
Why a password at all? Live calls consume OpenAI Realtime spend. The password gates only the live surfaces (calls that cost money). Everything read-only — the marketing site, the scripted
/demo, and the manager dashboard — is fully open, no password.
Before you start: headphones
Put on headphones before running any single-device live call. On one machine the platform simulates both the worker and the customer through one microphone — without headphones, the speaker feeds back into the mic and garbles the transcript. In a real two-person setup (see "Two-party live call" below) this never happens.
The scripted demo — /demo
(No password, $0 per view.) A self-contained, deterministic playback of a full live call. Nothing here calls OpenAI, so it's free to run any number of times and always behaves identically — ideal for a first look or a fallback if live audio has issues.
How to use:
- Open
/demo. - Click the large "▶ Play live demo" button over the transcript.
- Watch the scripted call play out: the voice twin, the D1 interrupt (customer cuts off the agent → agent audio stops + an "INTERRUPTED" badge flashes), the accent A/B beat, live coach tips, the sentiment graph, and the post-call review modal at the end.
- Controls at the bottom: Pause / Restart / Replay / mute.
Use it for: demos to anyone, screen recordings, sharing a link publicly. It needs no password and no microphone.
The worker console (live human agent) — /app
(Password on Start call.) The real thing: you speak as a support agent, and the customer hears a clean American-English voice twin of you in real time. A live coach watches the call and surfaces tip cards; a sentiment tracker graphs both speakers; at the end you get a post-call review with the scoring rubric and compliance section.
Running a call — three steps:
- Pick a customer. Choose one from the Customer dropdown. Their past-call memory loads automatically — watch the green memory: NNNc badge appear. You can also pick a voice twin (verse / marin / cedar) before starting.
- Click "Start call". Because this is a live call, a password dialog appears → enter the access password → the bridge connects and the call begins. The button turns into a red End call timer.
- Talk, and simulate the customer. Speak normally as the worker — the customer hears your neutralized twin. To play the customer side, press and hold the pink Customer mic (next section).
First-visit help: a "How to" walkthrough opens automatically; reopen it any time with the ? button.
The Customer mic — press & hold
The pink Customer mic button (bottom of the right column) is press-and-hold, like a walkie-talkie:
- Click and hold — the mic captures your voice as the customer for as long as you hold it. The label reads "Listening… release to send."
- Release — the moment you let go, the agent treats everything you just said as one finished customer turn.
This is what drives the live interrupt (D1): if you start holding the mic while the worker's voice twin is still talking, the twin audio cuts off instantly — exactly like a real customer talking over the agent. The customer side is passed through as raw audio (no twin) so it sounds like a separate person.
Reading the live signals
- Coach tips (right column) — high-priority suggestions surface every few seconds based on the rolling conversation. Mark one ✓ Used or ✕ Dismiss; they auto-clear after 30 seconds.
- Sentiment graph (under the transcript) — two lines track the worker (blue) and customer (purple) emotional intensity turn by turn. Hover a point for the model's reasoning.
- Live rubric (right column) — a 4-attribute projection of how the call is scoring. It stays hidden until a few turns in, because a short window is too noisy to score fairly.
Customer memory
Each customer has a memory built from their prior calls. When you select a customer, the top few relevant snippets are injected into the coach so its advice is grounded in history ("last call they were frustrated about a late delivery"). After every call, the new review is indexed back into that customer's memory, so the system gets smarter over time.
After the call
Click End call and a post-call review is generated: a sentiment arc, your strongest moments, concrete improvement areas, a full 7-attribute coaching rubric with evidence, the customer's key concerns, and the India-compliance section (per-rule pass/fail + verbatim evidence). Every call is saved:
/app/my-calls— your call history: date, customer, duration, CSAT, rubric, and click any row for the full review (including its compliance breakdown)./app/coaching— your coaching view: overall score, trend sparkline, your weakest attribute ("focus area") with what-to-do + example phrasing, and a card per rubric attribute.
The AI Agent Studio (live AI agent) — /agent
(Password on Start agent call.) Here you play the customer and talk to a live AI support bot (Atlas or Nova). When you end the call, the bot is scored on the same rubric as a human — landing on the hybrid leaderboard under that bot's name (agentType = 'ai'). This is the "grade the bot too" surface.
How to use:
- Open
/agent. - Pick a bot — Atlas (crisp, factual, low-empathy) or Nova (warm, but occasionally states things it isn't sure of).
- Click "Start agent call" → enter the access password when prompted → the bot connects.
- Speak as the customer. The bot responds live (open mic, server VAD — it auto-responds when you finish and can be interrupted).
- Live compliance nudge: if the bot doesn't disclose its identity early, a compliance nudge fires at ~22 seconds.
- Mid-call self-correction (Live Agent Corrector): if the bot misinforms (contradicts itself), skips a required compliance line, or sounds robotic, a supervisor steers it to correct itself out loud — e.g. "I'm sorry, I misinformed you earlier…" or "let me check that for you." (This runs by ear today — there's no visual panel for it yet.)
- End the call → it's scored and saved under the bot.
Two-party live call (console split) — /app + /call
(Needs two devices.) For a realistic call, put the agent on one device and the customer on a second device — no shared microphone, no acoustic feedback, clean transcripts on each side.
Device 1 — Agent (host):
- Open
/app→ unlock with the access password. - Use the "Go live · 2-party" control to create a room. The room now waits for a customer.
Device 2 — Customer (guest):
- Open
/call→ a password dialog opens automatically → enter the access password. - Pick a persona ("who are you on this call?").
- The agent's room appears in the room list → Join it. (The list also shows always-on bots Atlas / Nova — join one of those to be a customer talking to an AI agent instead.)
- Talk. The agent hears you; you hear the agent's voice twin.
Note:
/callis only the customer/guest side — it lists rooms and joins them. It does not start calls on its own. The agent side is always/app(human host) or a bot (AI host).
The manager dashboard — /manager-preview
(No password.) The hybrid-floor view. One leaderboard with both humans and AI agents, scored on one rubric.
What you'll see:
- Team stat tiles (workers, calls, team rubric average).
- Leaderboard — every agent (human + bot) with per-attribute scores, a trend sparkline, and overall.
- Compliance % column + a human-vs-AI split tile (who's more compliant).
- Click any row → drilldown with that agent's calls → click a call → the full review + compliance evidence.
- CSV export and an auditor compliance evidence export (per-rule, transcript-anchored).
Password reference
| Surface | Password prompt appears |
|---|---|
/app — Start call | On clicking "Start call" |
/agent — Start agent call | On clicking "Start agent call" |
/call — customer console | Automatically on page load |
/demo, /, /about, /manager-preview, /app/my-calls, /app/coaching | — (no password) |
Notes:
- The password is never stored in the app or the browser bundle — you type it at call time; it's held only as a server-side secret on the bridge. To get it for a live call, reach out via LinkedIn.
- The first live call may take ~2–5 seconds to connect while the bridge server wakes from idle. This is normal, not an error.
- Always use headphones for any single-device live call — the customer playback is what the agent hears, so open speakers cause feedback.
Troubleshooting
| Symptom | Cause / fix |
|---|---|
| The status reads "Bridge offline" and Start call is disabled | The WebSocket bridge isn't reachable. On the deployed site the first call wakes the bridge (~2–5s) — wait and retry. Running locally, start it with npm run dev:bridge (port 3001) and refresh. |
/call stuck on "Loading customers…" with no prompt | (Fixed 2026-07-01.) If it recurs, hard-refresh — the password dialog should auto-open on load. |
| "Start call" seems dead | It's clickable while locked — clicking it opens the password dialog. |
| Transcript shows garbled or doubled-up words | Acoustic feedback — the speaker is bleeding into the mic. Put on headphones. Goes away entirely in a real two-person setup. |
| The customer mic seems stuck on | It only stays live while held. If your pointer slid off the button mid-hold, click and release it once more — releasing always ends the turn. |
| Echo / feedback on a live call | You're on open speakers. Use headphones. |
Customer room doesn't appear on /call | The host must create the room first (on /app); the guest list polls every few seconds — give it a moment or hit refresh. |
| The console says it needs a wider screen | The live console requires a desktop browser at 1280px or wider. Resize the window or move to a larger screen. |
| Higher-than-expected voice latency | Two-party calls route through the hosted bridge (extra network hops); first call also pays cold-start. Expected on the deployed path. |
What's live vs. planned
(So the guide stays honest.)
- Live now:
/demo,/app(voice twin + coach + sentiment + review),/agent(live AI bot, scored),/calltwo-party,/manager-preview, India-compliance scoring (Layer 2), mid-call AI self-correction (Live Agent Corrector). - Built, not yet on the live site: Knowledge-Base RAG + domain guardrails (bind the agent + coach to an uploaded handbook, e.g. a Tesla manual).
- Planned: the NVIDIA-native migration (NeMo + Gnani Vachana for Indic/Hinglish), third-party bot-call ingestion, and multi-tenant auth. See the deck + the Why Voice Copilot page.