AWS Builder Center
Weekend Showcase Challenge: RELAY - Universal Autonomous AI Voice Operations Platform

Weekend Showcase Challenge: RELAY - Universal Autonomous AI Voice Operations Platform

RELAY is an autonomous AI voice platform that answers missed business calls, holds natural multilingual conversations grounded in a company's own knowledge base, and books confirmed appointments - built on AWS Amplify with Next.js 16 and Amazon Bedrock.

Building Tech for Impact - @TechSangi

Vision & The Problem It Solves

Every day, businesses lose revenue to missed phone calls. A buyer calls a real estate broker to book a property tour, a client calls an IT consultancy for an architecture review, a driver calls an automotive service hub after hours — and if nobody picks up, the deal often goes to whoever answers next. Traditional IVR phone trees ("Press 1 for sales, Press 2 for support") frustrate callers, and staffing phones around the clock is cost-prohibitive for most companies.
RELAY is an autonomous voice operations platform built to solve this for any business, regardless of size or industry.
RELAY connects PSTN phone lines, CRM systems, and a business's own knowledge base to conversational AI voice agents. When an incoming call goes unanswered, RELAY triggers an automated callback within seconds, holds a natural conversation in one of seven languages, answers questions using the business's own grounded FAQs, and books a confirmed calendar appointment — without a human ever picking up the phone.

How I Built It: Process, Key Decisions & Challenges

Building a telephony application that feels instant and trustworthy meant solving three hard problems: low-latency call dispatch, multilingual voice personas that get grammar right, and keeping the AI grounded so it never invents answers.
1. A non-blocking call-dispatch pipeline
A naive synchronous integration waits on the carrier's SIP handshake before responding, which can block a server thread for 1–3 seconds. Instead, I built an asynchronous dispatch pipeline in Next.js 16's App Router: the  /api/trigger-overflow  endpoint validates the phone number (E.164), maps it to the right regional carrier locale, and hands the call off to the CALL-E REST gateway — all before returning a response. The API itself responds in well under 20ms; the actual call setup and conversation happen asynchronously afterward. The frontend then polls call status every 300ms to show live handset events (Queued → Ringing → In-Progress → Completed) as they happen.
2. A consistent voice persona with correct Hindi grammatical agreement
Multilingual voice agents run into a subtle but important problem: grammatical gender. In Hindi and Nepali, a female-voiced agent needs to use feminine verb forms ("मैं बात कर रही हूँ", "मैं आपकी सहायता कर सकती हूँ") rather than the default masculine forms ("कर रहा हूँ"). I wrote an explicit persona prompt — warm, respectful, consistently feminine in its grammar — and applied it across every regional greeting so the voice feels coherent rather than machine-translated.
3. Regional carrier mapping and grounded knowledge to prevent hallucination
Different PSTN carrier gateways restrict which languages can open a call in a given region — for example, Indian (+91) numbers need a supported opening locale like  hi-IN  or  en-US . RELAY automatically detects the destination region and maps each call to a compliant opening locale, avoiding gateway rejections. Just as important: every business branch has its own editable knowledge base — specialists, pricing rules, FAQs pulled from the company's actual site — so the agent answers from real, operator-approved information instead of guessing.

AWS Services Used & Architecture Overview

RELAY runs as an event-driven system on AWS Amplify, organized into four layers:

Key AWS components:
  • AWS Amplify (Web Compute): Hosts the full Next.js 16 app with SSR, dynamic API routes, and CI/CD straight from GitHub.
  • AWS IAM & CloudWatch: A scoped IAM role gives Amplify's SSR compute permission to write logs, so I can watch webhook activity and API error rates in real time.
  • Amazon Bedrock: Runs post-call intelligence — extracting caller intent, drafting a follow-up SMS, and scoring call quality — after each conversation ends.

What I Learned Across the Summer

Three things stuck with me from this build:
  • Latency shapes trust in voice apps. Every added delay chips away at the illusion that you're talking to something responsive. Pushing heavy work into async handlers while keeping the UI optimistic made a bigger difference than any other single change.
  • Voice AI can't be one-size-fits-all across languages. Tone, respectful address, and grammatical gender agreement aren't cosmetic details — they're what makes a caller trust the agent instead of hanging up.
  • Amplify removes a lot of the deployment tax. Shipping a Next.js 16 App Router project with serverless SSR, environment isolation, and CDN distribution in one push meant I spent the weekend on the voice pipeline instead of infrastructure.

Links & Community Credit

Tag a Builder Who Inspired You

A big shoutout to  Hardik Bandhiya  SBGL, also a good friend of mine now. His Support to the AWS SBG Community and people is unmatched and the energy he brings is next level.  
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article