Skip to content
BranchesTags

Latest commit

 

History

2 Commits

Folders and files

NameName
Last commit message
Last commit date
config
config
 
 
docs
docs
 
 
prompts
prompts
 
 
scripts
scripts
 
 
web
web
 
 
.env.example
.env.example
 
 
.gitignore
.gitignore
 
 
README.md
README.md
 
 
package.json
package.json
 
 

Repository files navigation

Cold-Calling AI Voice Agent (Vantora · Alex)

Likva9 technical assessment — production-style cold-calling AI voice agent with live web demo, outbound phone calls, call recordings, transcripts, English/Urdu switching, and a clear path to full autopilot.
Status ✅ Live + source on GitHub
Live Demo https://web-rouge-xi-23.vercel.app
GitHub https://github.com/rohaan2802/Cold-Calling-Agent
Deadline 25 August 2026
Brand / Agent Vantora · Alex
Built by Rohaan
Email for submission contact@likva9.tech

Quick answers (read this first)

Do VAPI credits renew every day?

No.

Thing Renews daily? What it is
VAPI wallet credits (~$10) No — one-time when you create the account Spent on every web/phone call (STT + LLM + TTS + VAPI platform fee). When empty, calls stop until you top up.
Free VAPI US phone number — outbound limit Yes (~10 outbound attempts/day) Separate from credits. Even with money left, free numbers throttle outbound. Failed dials can count.
Groq Free tier (rate limits) LLM used via VAPI
Vercel Hobby Free hosting Live site stays up while project exists

Important: Paying $1–2 into VAPI wallet does not remove the free-number daily outbound cap. For more outbound, import Twilio / Telnyx / Vonage (or wait for the next day’s free-number reset).

Official refs: VAPI Pricing · VAPI Free Telephony · Outbound calling

Live link

https://web-rouge-xi-23.vercel.app

Stays live indefinitely while the Vercel project + env keys stay active and VAPI still has credits.

Source: https://github.com/rohaan2802/Cold-Calling-Agent (public)

Table of Contents

  1. Live link, uptime & evaluator guide
  2. Platforms & links (full stack)
  3. Credits, limits & costs (VAPI / Groq / Vercel)
  4. What is a cold-calling AI voice agent?
  5. What works today (shipped features)
  6. Architecture (step by step)
  7. Project map — which file does what
  8. Language switching (English / Urdu)
  9. Call History rules (what appears)
  10. Setup guide (step by step)
  11. Configuration & customization
  12. Deployment
  13. Improvements & how advanced it can get
  14. Full automation vision (autopilot)
  15. Testing checklist
  16. Troubleshooting
  17. Submission
  18. Changelog
  19. GitHub (source)

1. Live link, uptime & evaluator guide

Item Detail
Live Demo https://web-rouge-xi-23.vercel.app
Hosting Vercel (Hobby / free)
How long live? Indefinite while project + account stay active
Can stop if Project deleted, env vars removed, or VAPI credits exhausted

How to evaluate (recommended order)

  1. Open https://web-rouge-xi-23.vercel.app
  2. Select English or Urdu → Call Agent → allow microphone
  3. Talk naturally; try objections (“busy”, “send email”, “not interested”)
  4. End call → Call History → open a call → Play recording + read transcript
  5. Optional: Delete a test call
  6. Optional: Outbound to a US +1 number (subject to free-number daily limit)

Prefer Call Agent for evaluation — it uses credits but does not burn the free phone daily outbound quota.

Outbound test status (assessment note)

A full successful outbound phone test was not completed during this build window because the VAPI free US number hit the daily outbound call limit (org-level ~10 attempts/day; failed dials can count). Web Call Agent, recordings, transcripts, EN/Urdu, and inbound labeling were verified instead. Outbound should be re-tested after the free-number quota resets, or after importing Twilio / Telnyx / Vonage.

Language lock (all call types)

Once a language is locked for a call (web Call Agent, outbound, or inbound after language is set):

  • Agent replies only in that language.
  • If the user speaks another language, Alex must say exactly one lock line naming only the locked language:
    • English lock: Sorry, I can speak and understand only English.
    • Urdu lock (khalis Roman Urdu): Mazrat, main sirf Urdu hi bolta aur samajhta hoon.
  • Do not list multiple languages. After the lock line, if they speak the locked language again, continue the call normally.
  • Inbound language picker menu (offer EN in English / UR in Urdu at call start) is still planned below; until then inbound uses the assistant default (English) with the same lock script.

2. Platforms & links (full stack)

Platform Role Links
VAPI Voice orchestration, web calls, outbound, recordings, transcripts vapi.ai · Dashboard · Docs
Groq LLM (llama-3.3-70b-versatile) groq.com · Console
Deepgram Speech-to-text (via VAPI) deepgram.com · Docs
Vercel Live Next.js hosting vercel.com
Next.js 15 App UI + API routes nextjs.org
VAPI Voices (Elliot) Text-to-speech Via VAPI assistant voice config
VAPI Free US Number Outbound telephony (US-only on free) Dashboard → Phone Numbers

Optional / future telephony (not required for this demo)

Provider Why Link
Twilio Import number → bypass free-number daily outbound cap twilio.com
Telnyx Same idea; often trial credits telnyx.com
Vonage Import into VAPI vonage.com

2b. Free VAPI US number, TextMe testing & Twilio alternatives

What is the free VAPI US number for?

When you create a Free VAPI Number (e.g. +1 828 744 7146), that number is the caller ID / “from” line.

Question Answer
Calls go through this number? Yes — Alex dials out from this VAPI number
Calls arrive on this number for you to answer? Only if someone dials inbound to it (optional). For our demo we use it mainly for outbound
Where is it configured? VAPI Dashboard → Phone Numbers → copy Phone Number ID → .env as VAPI_PHONE_NUMBER_ID
Your TextMe number role? TextMe = the customer / “to” number Alex calls 
Alex (VAPI free number +1828…)  ──calls──►  Your TextMe (+1…)
     FROM / caller ID                         TO / prospect (you)

Fallback destination — what number to put?

For basic outbound testing you usually leave Fallback empty / unused.

Field When needed What to put
Fallback destination Only if you enable call transfer / human handoff and the transfer fails A real phone that should ring as backup (e.g. your TextMe +1…)
Outbound dialer on our site Always Prospect number = TextMe +1… in the form — not the VAPI number

Do not put the VAPI free number as the outbound “to” number — that would try to call itself.

How to test a free real call with TextMe (no Twilio)

  1. Install TextMe → get a US number (+1…)
  2. Keep TextMe open, notifications ON
  3. Open live site → Outbound Dialer
  4. Enter your TextMe as E.164, e.g. +12125551234
  5. Tap Call now
  6. Answer on TextMe — caller ID should look like the VAPI free number
  7. After call → Call History → play recording

Limits: free VAPI number ≈ 10 outbound attempts/day, US only. If you already hit the limit today, wait until the next day reset (or use a paid import number).

Credits: each call still spends one-time VAPI wallet credits (not a daily credit gift).

Twilio free alternatives (honest)

Option Free? Notes
VAPI free US number (what you have) Yes Best $0 path — daily outbound cap + US-only
Wait for daily reset Yes Same number, more attempts tomorrow
Telnyx trial Often free credits Buy/import number → VAPI Phone Numbers → Import (telnyx.com)
Vonage trial Trial credits common Import into VAPI
Twilio trial Often ~$15 trial credit Number ~$1/mo after; most documented import
Extra free VAPI numbers Yes Does not remove the org daily outbound cap

There is no unlimited free Twilio clone. Closest free path = VAPI free number + TextMe + wait for daily reset. Closest “pay almost nothing now” = Telnyx / Vonage / Twilio trial.

Import path (when you want to bypass daily free-number limit)

  1. Sign up Telnyx or Twilio → get a US number
  2. VAPI Dashboard → Phone Numbers → Import
  3. Put new Phone Number ID in .env / Vercel as VAPI_PHONE_NUMBER_ID
  4. Redeploy web
  5. Dial TextMe again from Outbound Dialer

3. Credits, limits & costs (VAPI / Groq / Vercel)

VAPI credits — not daily

Fact Detail
Starter credit About $10 one-time on new account signup
Daily renew? No
Monthly free allowance? No (usage-based after credits)
What spends credits Web Call Agent + outbound phone + STT/TTS/LLM pipeline
After credits empty Top up wallet / add payment method — otherwise voice calls fail

Free VAPI phone number — daily outbound limit

Fact Detail
Daily renew? Yes — outbound attempt quota resets per day
Typical cap ~10 outbound attempts/day per org on free VAPI numbers
Destinations US (+1) only on free numbers
Failed dials Can still count toward the daily limit
Does topping up credits remove this? No
How to scale outbound Import Twilio / Telnyx / Vonage number into VAPI

Other free pieces

Service Cost model
Groq Free API tier (rate limits apply)
Vercel Hobby Free hosting for this demo
Deepgram Billed through VAPI usage (covered by VAPI credits while they last)

4. What is a cold-calling AI voice agent?

An AI that talks like a human sales rep on a live call:

  1. Short intro + permission to continue
  2. Discovery questions
  3. Objection handling
  4. Ask for next step (demo / callback / email)
  5. Save recording + transcript for review

Alex represents Vantora (workflow automation: follow-ups, reminders, data entry — free trial then paid plan in the pitch).

Mode Who uses it Notes
Call Agent Recruiter / tester Browser mic; best for demos
Outbound Dialer Real phone Free number = US + daily cap
Call History Anyone Play / seek / delete recordings

5. What works today (shipped features)

Feature Status Notes
Live site on Vercel ✅ https://web-rouge-xi-23.vercel.app
Call Agent (web voice) ✅ Louder live playback boost; Elliot voice @ 0.8 speed
English / Urdu (choose before call) ✅ Mid-call live switch locked for now (roadmap §13)
Language lock scripts ✅ Wrong language → one lock line, then resume locked language
Idle presence (client-managed on web) ✅ ~5s → check → ~5s → check → ~5s → polite end (EN/UR match call language)
Hangup / “call cut” ✅ Roman + Urdu-script detection; Allah hafiz / goodbye → hard cut ~2s after
Callback ≠ hangup ✅ “call kr do / call me back” books callback; does not end the call
Soft barge-in ✅ Needs several real words before Alex stops mid-line (avoids half-salam cuts)
Outbound dialer ✅ US +1 on free number (daily cap applies)
Call History ✅ Only calls with recordings; paginated past deleted stubs
Recording play + seek (±10s) ✅ Blob load + range proxy
Transcripts ✅ Per call detail
Delete call ✅ VAPI delete + cookie hide so refresh doesn’t bring it back
Objection playbook ✅ prompts/objections.md
Groq via VAPI ✅ llama-3.3-70b-versatile · short maxTokens for snappy replies
Dark responsive UI ✅ Phones / tablets / desktop
Favicon / branding ✅ Vantora
Autopilot (auto leads + auto dial) 🔜 Designed in §14 — not fully coded yet

Assessment completeness (honest)

Area Status
Web Call Agent + EN/Urdu + history + recordings Done — primary eval path
Hangup after goodbye / “call cut” Shipped — re-test after hard refresh
Outbound full E2E on free VAPI number Pending daily quota / re-test after reset or imported number
Live mid-call language switch + inbound IVR Roadmap only (§13)
Callback scheduler / Excel / dashboard / autopilot Roadmap only (§13–14)

6. Architecture (step by step)

Step-by-step call flow (Call Agent)

  1. User opens live site → selects language → Call Agent
  2. Browser loads @vapi-ai/web with public key
  3. vapi.start(assistantId, overrides) applies EN/Urdu prompt, voice, STT
  4. Mic audio → Deepgram STT → Groq LLM → Elliot TTS → speaker (volume boosted in browser)
  5. Call ends → VAPI stores recording + transcript
  6. UI lists call via GET /api/calls → detail plays via /api/calls/[id]/recording

Step-by-step outbound flow

  1. User enters +1… in Outbound Dialer
  2. POST /api/outbound → VAPI /call with phoneNumberId + assistantId
  3. Free US number dials prospect (if daily limit allows)
  4. Same agent pipeline on the phone
  5. Recording appears in Call History when artifacts exist

Diagram

┌─────────────────────────────────────────────────────────────────┐
│  Live Web App — Vercel                                           │
│  Call Agent · Outbound Dialer · Call History · Call Detail       │
└───────────────┬─────────────────────────────┬───────────────────┘
                │ @vapi-ai/web                │ Next.js /api/*
                ▼                             ▼
┌──────────────────────────────┐   ┌──────────────────────────────┐
│ VAPI Web Session             │   │ API: calls, outbound,        │
│ Deepgram → Groq → Elliot     │   │ recording proxy, webhook     │
└──────────────────────────────┘   └──────────────┬───────────────┘
                                                  │
                                                  ▼
                                   ┌──────────────────────────────┐
                                   │ VAPI Cloud                    │
                                   │ Assistants · Calls · Storage  │
                                   │ Free US Phone Number          │
                                   └──────────────────────────────┘

Data ownership

Data Where
Prompts / personality prompts/ → pushed with scripts/deploy-assistant.js
Voice / model / recording flags config/assistant-config.json
Call audio + transcripts VAPI cloud
Hidden deleted IDs Browser cookie vantora_hidden_calls (so refresh stays clean)
Web UI web/ on Vercel
Secrets Root .env + Vercel Environment Variables

7. Project map — which file does what

Root

Path Responsibility
README.md This documentation
.env / .env.example API keys (never commit real secrets)
package.json dev, deploy:assistant, sync:web-env, add:groq
config/assistant-config.json Default VAPI assistant: voice, model, recording, timeouts
prompts/system-prompt.md English base personality + pitch + flow
prompts/objections.md Objection → reply playbook
scripts/deploy-assistant.js Push prompts + config to VAPI
scripts/add-groq-credential.js Attach Groq key to VAPI
scripts/sync-web-env.js Root .env → web/.env.local
scripts/run-dev.cmd Windows: free port + Next + Chrome
.vscode/* IDE Node path + launch

Web pages (web/app/)

Path Responsibility
layout.tsx Shell, nav, fonts, viewport, favicon
page.tsx Home: hero + Call Agent + Outbound + History
globals.css Theme, responsive layout, tap-highlight off
calls/page.tsx Full Call History page
calls/[id]/page.tsx Single call detail route
icon.svg Favicon

UI components (web/components/)

Path Responsibility
TalkToAgent.tsx Call Agent, EN/Urdu lock, idle, hangup timers, live volume boost, transcript
OutboundDialer.tsx Place real outbound call
CallHistory.tsx List + delete
CallDetail.tsx Meta, player, transcript, delete
AmplifiedPlayer.tsx Recording play / seek / loudness

API (web/app/api/)

Path Responsibility
calls/route.ts List calls (recording-only; hide deleted / cookie-hidden)
calls/[id]/route.ts Get one call · DELETE (+ Set-Cookie hide)
calls/[id]/recording/route.ts Proxy audio for play + seek
outbound/route.ts Start outbound VAPI call
webhook/route.ts VAPI event receiver

Libs / hooks

Path Responsibility
web/lib/vapi.ts Types, labels, history filters, cookie helpers, vapiFetch
web/lib/speechConfig.ts STT (Deepgram), voice speed, idle lines, hangup/goodbye detectors, end-call phrases, turn-taking
web/lib/englishSystemPrompt.ts English Call Agent system prompt (web overrides)
web/lib/urduSystemPrompt.ts Roman Urdu system prompt + callback-vs-cut rules
web/lib/languageLock.ts EN/UR lock one-liners when user speaks the wrong language
web/lib/callErrors.ts Friendly error strings for UI
web/scripts/patch-assistant-efficient.js Patch base VAPI assistant (e.g. clear server idle hooks for web)
web/hooks/useAutoDismiss.ts Auto-clear UI messages (~5s)
web/app/icon.svg Favicon / browser tab icon (Vantora V + voice-wave) — Next.js serves this automatically

Change X → edit Y

Want to change Edit Then
English pitch / personality (base) prompts/system-prompt.md npm run deploy:assistant
English Call Agent prompt (web) web/lib/englishSystemPrompt.ts redeploy web
Objections prompts/objections.md deploy assistant
Default voice/model/recording config/assistant-config.json deploy assistant
Urdu wording / cut vs callback web/lib/urduSystemPrompt.ts redeploy web
Hangup phrases / idle lines / STT web/lib/speechConfig.ts redeploy web
Call Agent / volume / hangup timers TalkToAgent.tsx redeploy web
History filters web/lib/vapi.ts + api/calls redeploy web
Theme / responsive globals.css redeploy web

8. Language switching (English / Urdu)

Control Behavior today
Before call (web) User picks English or Urdu, then starts Call Agent — language is locked for that call
During call (web) Language buttons are disabled — mid-call live switch is not available right now
Outbound / inbound No live language switch yet (English assistant default on phone unless configured otherwise)
English mode Speaks and understands English only (language lock)
Urdu mode Speaks and understands Urdu (replies in easy Roman Urdu on screen)

Future: full live language switching for web + inbound + outbound (digit IVR at start + mid-call digit switch, plus efficient web switch) — see §13.

Files: TalkToAgent.tsx, urduSystemPrompt.ts, speechConfig.ts, base rules in prompts/system-prompt.md.

9. Call History rules (what appears)

Appears in history? Condition
Yes Call has a saved recording artifact
No Soft-deleted (call-deleted) or user Delete (cookie hide)
No Failed outbound / daily-limit / error (usually no recording)
Yes (often) Alex spoke, user hung up without talking — if VAPI still saved audio

Labels: Web demo N or Outbound (phone), never raw outboundPhoneCall on browser demos.

10. Setup guide (step by step)

Accounts

# Platform Cost Link
1 VAPI ~$10 one-time credits https://dashboard.vapi.ai
2 Groq Free https://console.groq.com/keys
3 Vercel Free Hobby https://vercel.com

Steps already done for this project

  1. VAPI account + assistant
  2. Groq key → npm run add:groq
  3. Deploy assistant → npm run deploy:assistant
  4. Free US phone number → VAPI_PHONE_NUMBER_ID
  5. Next.js web app + Vercel env + prod deploy
  6. README + live link

Env vars

VAPI_API_KEY=
VAPI_PUBLIC_KEY=
VAPI_ASSISTANT_ID=
VAPI_PHONE_NUMBER_ID=
GROQ_API_KEY=
NEXT_PUBLIC_VAPI_PUBLIC_KEY=
NEXT_PUBLIC_VAPI_ASSISTANT_ID=

Sync local web: npm run sync:web-env

Local run (Windows)

cd "C:\Users\CodeTech\Desktop\Cold-Calling AI Voice Agent"
.\scripts\run-dev.cmd

Or npm run dev — picks a free port from 3000 and opens Chrome.

11. Configuration & customization

Change File Command
Pitch / product prompts/system-prompt.md npm run deploy:assistant
Objections prompts/objections.md npm run deploy:assistant
Voice / model / timeouts config/assistant-config.json npm run deploy:assistant
Web UI / APIs web/** cd web → npx vercel --prod --yes
Secrets .env + Vercel env Redeploy web after NEXT_PUBLIC_* changes

Default stack: Groq llama-3.3-70b-versatile · VAPI voice Elliot · recording ON.

12. Deployment

Component Platform URL / place
Web demo Vercel https://web-rouge-xi-23.vercel.app
Assistant VAPI Dashboard → Assistants
Phone VAPI Free Telephony Dashboard → Phone Numbers 
npm run deploy:assistant
cd web
npx vercel --prod --yes

13. Improvements & how advanced it can get

These are future upgrades (not all built yet). The live demo already does Call Agent, outbound, recordings, EN/Urdu, idle, hangup. Below is how far this product can go — including efficiency techniques and near 100% hands-off automation.

Already shipped (today)

  • Live Call Agent + Outbound dialer + Call History + recordings
  • English / Urdu · delete · friendly errors · louder live voice · Elliot @ 0.8 speed
  • Language before call (web): user chooses English or Urdu, then starts — that language stays locked for the whole call. Mid-call live switch is intentionally disabled for now (see roadmap below).
  • Language lock script on web / outbound / inbound (once language is set): wrong language → English: Sorry, I can speak and understand only English. · Urdu: Mazrat, main sirf Urdu hi bolta aur samajhta hoon. (then resume when they use the locked language)
  • Idle presence (5s → 5s → end): client-managed on web so Urdu never hears English “Are you there?”; phone/outbound can still use VAPI hooks when not client-managed
  • Hangup reliability: “call cut / call kt / کال کٹ …” → short bye → hard stop ~2s after Allah hafiz / goodbye (client timer + end-call / stop, not only LLM hope)
  • Callback ≠ cut: “call kr do / call me back” books a time; does not hang up
  • Soft barge-in (stopSpeakingPlan) so ambient noise doesn’t chop Assalam mid-line
  • Delete removes the call from the site and permanently deletes it on VAPI (recording included)

Nothing critical left for the assessment demo?

For Likva9 eval via Call Agent, the core loop is done. Remaining gaps are either ops limits or roadmap product:

Still open Why it isn’t a blocker for web demo
Outbound E2E on free VAPI number Daily ~10-attempt cap (re-test after reset / import Twilio etc.)
Mid-call live language switch Intentionally locked; design is in this section
Callback auto-dial / Excel / dashboard Product roadmap — agent can agree verbally but no scheduler yet
GitHub remote ✅ Done — https://github.com/rohaan2802/Cold-Calling-Agent

Efficiency techniques (latency · cost · reliability)

Ways to make Alex faster, cheaper, and more stable without changing the product story:

Technique Why it helps Effort
Keep short maxTokens (already ~110) + tight prompts Fewer tokens = lower Groq/VAPI cost + snappier TTS Low
Prefer one STT language per call (already locked) Multi-language STT is slower / noisier; dedicated en / ur is cleaner Done
Tune endpointing (Deepgram 300ms + start/stop speaking plans) Fewer false barge-ins and fewer “I didn’t hear you” loops Done / tweak
Client hangup hard-stop after goodbye (already) Don’t rely only on LLM endCall — STT miss / tool miss still cuts the line Done
Unit-test hangup / goodbye detectors (isUserHangupRequest, isAssistantGoodbye) Catch STT spelling variants (“all kt”, Arabic script) before prod breaks again Low
Webhook → own DB (Postgres / SQLite) instead of only listing VAPI calls Faster history, outcome labels, no pagination fights with call-deleted stubs Medium
Outcome classifier post-call (interested / callback / DNC) via one cheap LLM pass Dashboard + sheet sync without listening to every recording Medium
Optional cheaper model for simple turns; keep 70B for objections Cut LLM spend on “haan / theek hai” acknowledgements Medium
Prompt / FAQ RAG over product docs Shorter system prompt + accurate answers → less hallucination, fewer tokens Medium
Recording retention policy (auto-delete after N days) Lower storage / privacy risk Low–Med
Import Twilio/Telnyx + quiet-hours dialer Real outbound scale; free VAPI number stays demo-only Medium
Parallel dial / queue with concurrency caps More connects per hour once telephony is paid High
Pre-warm VAPI session / measure time-to-first-audio Demo feel = “human speed”; track p50/p95 Medium
CI on GitHub (lint + typecheck + hangup unit tests) Safer deploys; fewer “forgot to redeploy web” mistakes Low

Callback reminders — not built yet

We have not implemented automatic callback reminders yet. The agent can agree to call back in conversation, but there is no scheduler that dials again at a set time. That belongs in the improvements below.

Planned improvements (product roadmap)

Idea What it would do
Live language switching (web + inbound + outbound) Keep the option for all three call types: choose language at call start and switch mid-call. See detailed design below. Not enabled in the live demo right now.
Inbound language picker (EN / Urdu) When someone calls in, Alex first offers a choice (spoken + digit dial) — then locks that language. Not built yet.
Callback reminders If prospect says "call me at 5pm / kal", AI schedules reminder + auto-dials at that time
Auto-discard not interested Outcome = not interested / DNC → drop from active queue (keep archive for compliance)
Excel / CSV upload button Upload lead sheet → save → AI works that list
AI updates the sheet After each call, mark rows: Interested / Not / Callback / Email / Wrong number
Progress dashboard Totals: leads, called, remaining, interested %, callbacks due, calls by hour, agent progress
Client intake Website form / WhatsApp / email → leads enter queue automatically
Agent shares company details Tools: SMS/email with pricing, calendar link, brochure — without a human on the line
Hangup edge-case hardening More STT misspellings, dual-channel endCall + client stop, optional “End call” UI button during web demo
Own call DB + analytics Store outcomes, talk-time, connect rate — don’t only depend on VAPI list API
Full autopilot loop Dial → talk → log → remind → re-dial → nurture → book demo

Live language switching — full design (future)

Apply the same pattern to web Call Agent, inbound phone, and outbound phone:

1) At call start (phone IVR style)

  • Agent greets and offers language by pressing a digit, the way normal phone menus work (example: “For English, press 1. Urdu ke liye 2 dabayein.” — English line in English, Urdu line in Urdu).
  • After the digit, lock STT + replies to that language for the call.
  • Also tell the caller: during the call you can switch language by pressing a specific digit (example: “Anytime, press 9 to switch language.”).

2) Mid-call switch (phone)

  • User presses the announced digit → language switches live (prompt + speech recognition) without hanging up.
  • Agent confirms briefly in the new language, then continues the sales conversation.

3) Mid-call switch (web)

  • Same idea via UI (English / Urdu buttons) — must be reliable: if the user switches while Alex is talking, Alex should cut over and continue in the new language; transcript/STT must match the new language (no wrong-script / wrong-language replies).

4) Quiet / timing edge cases (all channels)

  • If both are quiet and the user switches → Alex should speak soon in the switched language (short confirm, then continue).
  • If the user is speaking and Alex is quiet, and the user switches → Alex should respond immediately, naming the switched language, e.g. English: “I understand only English.” / Urdu: “Main sirf Urdu hi samajhta aur bolta hoon.” (then continue in that language).

Status today: web allows language only before start; mid-call web switch is locked off. Inbound/outbound live switch + digit IVR are not implemented yet.

Level A — polish (easy)

  • Live language switching (web + inbound + outbound) per §13 design — digit IVR at start, mid-call digit / UI switch, quiet-edge confirms
  • Inbound IVR language menu (English prompt in English + Urdu prompt in Urdu → lock agent language for that call)
  • Stronger outcome labels (booked / callback / lost / not interested)
  • CSV / Excel export of call history
  • Unit tests for hangup / goodbye phrase detectors
  • Optional on-screen End call button (web) as backup to voice “call cut”
  • More TTS voices / A/B openings
  • Homepage "How to test" 3-step box
  • GitHub repo + CI (lint / typecheck)

Level B — sales ops (medium)

  • Callback reminder engine (cron + timezone + quiet hours)
  • Excel/CSV upload UI + editable lead table
  • AI sheet sync (interested / not / callback columns)
  • Analytics dashboard (funnel, connect rate, talk time, outcomes)
  • Webhook → own DB (faster history, no deleted-stub pagination pain)
  • Calendar booking (Cal.com / Google)
  • Email / SMS follow-up tools
  • Warm transfer to human only when requested
  • Cheaper model path for short acknowledgements

Level C — enterprise / near–no human (advanced)

  • Parallel dialing + smart queue
  • Lead discovery APIs with consent rules
  • Brand voice cloning + RAG over product docs
  • Compliance (DNC, quiet hours, recording consent + retention)
  • Multi-tenant SaaS (many clients, each with agent + dashboard)
  • Slack / email digests: "today 40 calls, 6 interested, 3 demos booked"

Ceiling: sheet in → AI dials → updates sheet → reminders → books meetings → reports — humans mostly set rules and top up credits.

14. Full automation vision (autopilot — without constant human intervention)

Goal: Keep the app live-deployed; the system works leads, calls, logs, reminds, and surfaces qualified clients — you mostly keep hosting + credits + policies.

Target loop

Client / lead intake (form, Excel upload, CRM, WhatsApp)
        → Lead DB (pending / calling / interested / not / callback / DNC)
        → Scheduler (Vercel Cron) — calling hours + callback times
        → Outbound API (existing /api/outbound)
        → Alex talks (EN/Urdu) + tools (email, SMS, calendar)
        → Webhook writes outcome back to sheet/DB
        → Dashboard updates live
        → Callbacks auto-fire later
        → Interested leads → notify owner + book demo

How new clients / prospects can contact (suggested)

Channel Flow
Website Call Agent Already live for demos
"Request a call" form Name, phone, company → queue
Excel upload (sales team) Bulk B2B lists
Inbound call to VAPI number Same agent answers inbound
WhatsApp / email webhook Parse → create lead

How the agent can give your details (suggested tools)

  • Email tool: pricing + calendar link
  • SMS tool: short follow-up with booking URL
  • Speak aloud: company name, offer, next step
  • Post-call summary to your inbox for interested leads only

What stays human (honest)

Humans still do System does alone
Keep Vercel + keys live Dial queue + callbacks
Top up VAPI credits Talk, record, label outcomes
Set DNC / quiet hours / caps Update Excel/DB + dashboard
Close high-value deals Notify you of interested / booked demos

When to build: after assessment acceptance, with credits + lawful leads + compliance.
Free-tier honesty: free VAPI number still has daily outbound caps + US-only until Twilio/Telnyx/Vonage.

15. Testing checklist

  • Live demo loads
  • Call Agent + first message
  • English / Urdu (before call only; mid-call switch locked)
  • Objections
  • Idle presence in correct language (EN / UR)
  • Hangup: “call cut kr do” / goodbye → cut ~2s after Allah hafiz
  • Callback phrase does not hang up (“call kr do” / “call me back”)
  • History + recording play + delete
  • Language lock script (EN lock in English; UR lock in Roman Urdu; then resume locked language)
  • Outbound US end-to-end — not completed in this window due to VAPI free-number daily outbound limit (re-test after reset or imported Twilio/Telnyx/Vonage)
  • Live language switching web + inbound + outbound (planned — see §13)
  • Inbound language picker / digit IVR (planned — see §13)

16. Troubleshooting

Problem Fix
Outbound "daily limit" Wait for next day reset, or import Twilio/Telnyx/Vonage
International dial fails Free number = US only (+1)
Credits gone / calls fail VAPI wallet one-time — top up (credits do not renew daily)
Recording missing Wait 30–60s, refresh; only recorded calls list
Deleted call returns Hard refresh; delete cookie + filter call-deleted
Soft Alex voice Live boost in TalkToAgent; raise device volume
“Call cut” doesn’t hang up Hard refresh live site; say clearly “call cut kr do”; client must detect hangup (see speechConfig.ts)
Stays open after Allah hafiz Should hard-stop ~2s after goodbye; if not, hard refresh — check TalkToAgent hangup timers
English “Are you there?” on Urdu call Web uses client-managed idle; patch base assistant hooks empty (patch-assistant-efficient.js)
Node not found in Cursor Use scripts/run-dev.cmd / .vscode Node path
Favicon / icon Source file: web/app/icon.svg (browser tab icon)

17. Submission
Live https://web-rouge-xi-23.vercel.app
GitHub https://github.com/rohaan2802/Cold-Calling-Agent
Platforms VAPI · Groq · Deepgram · Next.js · Vercel
Email contact@likva9.tech
Deadline 25 August 2026

Suggested note: Call Agent for eval · Call History for recordings · README covers callbacks, Excel, dashboard, efficiency techniques, and hands-off automation path.

18. Changelog

Date Change
2026-08-20 Project init — VAPI + Groq + Vercel
2026-08-20 Call Agent, outbound, history, recordings, delete
2026-08-20 Vantora brand, EN/Urdu, responsive UI
2026-08-20 Urdu nova-3 STT, louder live voice, seekable player
2026-08-20 History = recordings only; delete cookie hide
2026-08-20 README: credits one-time not daily, architecture, platforms, autopilot
2026-08-21 README: callbacks, Excel upload, discard, dashboard, client intake, hands-off automation; icon path documented
2026-08-21+ Language lock before/during call; client-managed idle; soft barge-in; voice speed 0.8
2026-08-23 Hangup: Urdu “call cut” + Allah hafiz → hard cut ~2s; README §5/§7/§13 efficiency + completeness
2026-08-23 Public GitHub: https://github.com/rohaan2802/Cold-Calling-Agent

19. GitHub (source)
Repo https://github.com/rohaan2802/Cold-Calling-Agent
Visibility Public
Branch main
Secrets .env / real keys are not in the repo (see .gitignore + .env.example)

Clone:

git clone https://github.com/rohaan2802/Cold-Calling-Agent.git
cd Cold-Calling-Agent
cp .env.example .env   # fill keys
npm run sync:web-env
cd web && npm install && npm run dev

Folder tree (source)

Cold-Calling AI Voice Agent/
├── README.md
├── .env.example
├── package.json
├── config/assistant-config.json
├── prompts/system-prompt.md
├── prompts/objections.md
├── scripts/
│   ├── deploy-assistant.js
│   ├── add-groq-credential.js
│   ├── sync-web-env.js
│   └── run-dev.cmd
└── web/
    ├── app/
    │   ├── icon.svg          ← browser favicon
    │   ├── page.tsx
    │   ├── layout.tsx
    │   └── api/…
    ├── components/
    ├── lib/
    │   ├── speechConfig.ts   ← STT, idle, hangup detectors
    │   ├── englishSystemPrompt.ts
    │   ├── urduSystemPrompt.ts
    │   └── …
    └── hooks/

Built by Rohaan — Likva9 Technical Assessment

About

Vantora · Alex — Cold-Calling AI Voice Agent (Likva9 assessment). Live: https://web-rouge-xi-23.vercel.app

web-rouge-xi-23.vercel.app

Topics

cold-callingnextjsvapivoice-ai

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages