DEV Community: Amrita Shrivastava

Building a Voice AI Agent That Turns User Drop-Offs Into Conversations

Amrita Shrivastava — Fri, 23 Oct 2026 13:29:00 +0000

Our team, Rho secured 2nd place at a Voice AI Buildathon by building an AI voice agent for user reactivation and referrals.

The most exciting part? We didn't just present a working prototype. We demonstrated it through a real outbound call, with Bachatt co-founders Mayank Agarwal and Ankur Jhavery, along with judges from Urban Company and OLX, interacting with our agent.

Here's what we built, the problem we wanted to solve, and how we approached it.

The problem: Analytics can tell you where users drop off, but not always why

Imagine a user downloads a savings app, starts onboarding, and abandons the process halfway through.

Your analytics dashboard can tell you where the user dropped off. It might show that they never completed KYC, linked their bank account, or made their first investment.

But it can't necessarily tell you what happened on the other side of the screen.

Did the user get confused?
Did they run into a technical issue?
Did they have concerns they couldn't resolve?
Did they simply need a little guidance?

Traditional re-engagement often relies on push notifications, emails, and automated reminders. These can bring users back, but they don't always help identify the underlying friction.

We wanted to explore a different approach.

What if, instead of sending another reminder, the product could actually call the user, understand their concern, and help them take the next step?

That became the core idea behind our project.

The solution: An AI voice agent for proactive reactivation

We built a voice agent around Bachatt's reactivation and referral journey. The broader concept, however, applies to any product where users encounter friction during onboarding or fail to complete an important action.

The agent focuses on two scenarios.

1. Reactivating users who get stuck

When a user stalls during a critical onboarding step, the system can identify the point of friction and initiate a relevant interaction.

Rather than delivering a generic message, the agent aims to understand the user's specific concern and provide contextual guidance.

For example, a user who hasn't completed a step should receive help relevant to that step, rather than a generic reminder to finish onboarding.

2. Helping referred users complete onboarding

Referral links bring prospective users into a product, but downloading an app doesn't guarantee that they will complete registration or become active users.

Our second scenario focuses on referred users who don't progress after receiving a WhatsApp referral link.

The proposed flow gives the user time to download and onboard. If they haven't progressed after a minimum 36-hour wait, the agent can call them, answer questions, help resolve minor issues, and guide them towards the next step.

The objective is not simply to get someone on a call. It's to remove the friction preventing them from moving forward.

How the system works

The central design principle was context-aware interaction.

A voice agent needs more than a language model and a phone connection to be useful in a real product workflow. It needs to know where the user is in their journey and what action makes sense next.

Our proposed system follows this high-level flow:

Detect: Identify a stalled onboarding journey or an incomplete referral journey.
Retrieve context: Query the user's current journey stage so the agent knows where they left off.
Initiate a conversation: Trigger an outbound call when the relevant conditions are met.
Understand the concern: Let the user explain what is preventing them from progressing.
Guide the next step: Provide relevant assistance within predefined product boundaries.
Feed insights back: Aggregate recurring friction points to help product teams understand why users struggle.

The system uses Model Context Protocol (MCP) to query live journey-stage information. This helps ground the interaction in the user's current state rather than relying on a generic script.

The onboarding journey can include stages such as authentication, KYC, profile setup, bank account linking, plan selection, and the first savings action.

The guiding principle is simple: help the user with the next relevant step, not the entire journey at once.

Why context matters in Voice AI

A conversational agent can sound natural and still be ineffective if it doesn't understand what the user needs.

Consider two users who haven't completed onboarding.

One might be stuck because they don't understand a particular step. The other might have encountered a technical issue.

Sending both users the same message ignores the difference between their situations.

By connecting the agent to the product's journey state, we explored how voice interactions could become more targeted and useful.

There is also an important boundary here, particularly in financial products. Our design keeps the agent within the scope of platform-related assistance rather than providing broad financial advice.

The goal is to help users navigate the product, not make decisions on their behalf.

The demo: Taking it beyond slides

One of the highlights of the buildathon was demonstrating the system through a real outbound call.

Bachatt co-founders Mayank Agarwal and Ankur Jhavery interacted with our agent, as did Parth Mahajan from Urban Company and Ajay Singh Rathore from OLX.

Getting to see people interact with the agent in real time made the experience very different from presenting an architecture diagram or a recorded demo.

It also opened up questions about what makes a voice agent genuinely useful: how it handles real conversations, how much context it needs, and how to ensure that its responses remain relevant to the user's situation.

We walked away with 2nd place, but the bigger takeaway was seeing the idea move from a concept to a working, demonstrable interaction.

What we're taking away from the build

This project reinforced a distinction that I think will matter increasingly as AI agents become part of everyday products.

There's a difference between an AI system that can hold a conversation and one that can take meaningful action within a product workflow.

The latter requires more than conversational ability. It needs context, access to the right information, clear boundaries, and a reason to initiate the interaction in the first place.

It also raises interesting engineering questions around interruption handling, latency, call reliability, user consent, escalation paths, and measuring whether these conversations actually improve outcomes. These are areas that any production-ready implementation would need to address carefully.

For me, the most interesting possibility is the feedback loop. Voice AI could potentially do more than help individual users complete a task. By surfacing recurring questions and points of confusion, it could help product teams improve the underlying experience for everyone.

That turns reactivation from a purely growth-oriented activity into a potential source of product insight.

A final thought

Voice is one of the earliest ways humans learn to interact with the world, long before we learn to write. It is an incredibly natural bridge between minds.

As we build increasingly capable AI systems, voice offers an opportunity to make technology feel less like a sequence of interfaces and more like an interaction.

We're only beginning to explore what that looks like in real products.

And congratulations to everyone who participated!

Loopcall: Meet Sonali, a Voice - First AI Agent that Confirms D2C Orders At Zero Cost.

Amrita Shrivastava — Mon, 19 Oct 2026 07:25:46 +0000

What I Built

Meet Loopcall, powered by Sonali, an AI voice agent built to solve a very real problem for small e-commerce businesses.

Incomplete delivery details mean more manual calls, delayed fulfilment, and wasted operational time. But there's another cost that often goes unnoticed: fake COD (Cash On Delivery) orders. When an illegitimate order gets processed, sellers can lose money on shipping, packaging, and failed deliveries.

I wanted to build a solution that tackles both problems without spending a dime on an expensive automation stack.

Sonali reads order details from a local Excel workbook, identifies the flagged concern, and calls the customer to resolve it through a natural, conversational voice interaction. Whether it's a missing pincode or confirming that a COD order is genuine, she focuses on asking the right questions and capturing the response.

But I didn't want to build another voice bot that talks endlessly and leaves the actual work to someone else.

Loopcall closes the loop. Sonali analyses the outcome, marks an order as verified only when the necessary checks succeed, updates the relevant fields in the original workbook, and flags complicated or ambiguous cases for human review.

The system is designed with practical guardrails: concise conversations, controlled processing, minimal unnecessary computation, and human intervention when the agent cannot confidently resolve an issue.

The other important decision was keeping the order data local. Instead of depending on Google Sheets authentication, spreadsheet APIs, or a cloud-hosted database, Loopcall uses an offline Excel workbook as its source of truth.

The goal is to combine zero-cost tooling, local-first data handling, open-source AI, and real-world task completion into something a small business can actually use.

Sonali doesn't just make the call. She helps determine whether the order is legitimate, resolves missing information where possible, and ensures the outcome is recorded.

That's Loopcall: fewer manual calls, fewer avoidable failed deliveries, and a verification process that doesn't stop at the conversation.

Demo

Link : Loopcall - Demo

The demo shows Sonali handling an order verification from start to finish:

Pick an order: Loopcall reads an incomplete or flagged order from the local Excel workbook.

Understand the issue: Sonali identifies the customer's details and the concern that needs to be resolved, such as a missing pincode and an unconfirmed/ unverified order.

Make the call: Sonali initiates an outbound call and guides the customer through a short, contextual conversation.

Evaluate the response: Based on the conversation, she determines whether the required checks are complete or the case needs human review.

Close the loop: She then updates the relevant fields in the original Excel workbook, preserving the order record and saving the outcome.

The important part isn't just hearing an AI agent speak. It's seeing the result reflected in the actual order data. From a flagged concern to a recorded outcome, Loopcall connects the conversation to the action that follows.

Code

GitHub repository: Loopcall

Loopcall is built around a complete order-verification workflow, connecting order selection, concern-specific dialogue, outbound calling, response analysis, and persistent status updates.

The local Excel workbook serves as the source of truth for the demo. Only the fields required for verification are updated, while unrelated order data remains unchanged.

How I Built It

Three open-weight models on one laptop:

Hearing — Whisper-Hindi2Hinglish-Swift, because Indian customers mix languages.
Thinking — Gemma 4 E4B via Ollama, for messy replies the rules cannot place.
Speaking — Kokoro-82M, Hindi voice, same person in every language
Plain rules own the script and the address checks. Gemma only steps in when the customer hedges, asks a question, or goes off-script. That keeps the call short and the bill at zero.

One of my key decisions was to use a local Excel workbook instead of Google Sheets, since that is exactly how the business (Sustainable Universe) also actually maintains their order inventory.

This also removes the need for Google sign-in, spreadsheet API credentials, and a cloud-hosted spreadsheet database. Sonali reads the relevant order details locally, while Loopcall writes the verification outcome back to the same workbook.

The workflow looks like this:

Excel Workbook → Order Selection → Sonali → Outbound Call → Response Analysis → Excel Update

The flagged concern determines what Sonali asks. A missing pincode, for example, should trigger a request to confirm the pincode rather than a generic verification script.

The system also distinguishes a successful verification from a failed or unanswered call. An attempted call alone is not enough to mark an order as verified.

Why Does Open Innovation Matter?

A hosted voice API would charge per minute and upload names, phones, and addresses. A small business making dozens of confirmation calls a day cannot afford that, and should not ship that data out.

Open models meant I could run the whole agent offline, swap Whisper when it failed on Hinglish, and keep Gemma as a helper instead of the whole product. That is what makes Loopcall a side hustle instead of a demo that dies on the invoice.

For Loopcall, open innovation matters in three ways.

Data ownership: Keeping order records in a local workbook gives the business direct control over its primary order data.

Flexibility: Open-weight models and open-source tools can offer more flexibility to customize, evaluate, and adapt AI systems to specific business workflows, subject to their licenses and capabilities.

Practical automation: The objective is not to build an impressive chatbot that stops at generating a response. It is to build an agent that performs a useful task, records the outcome, and knows when a human needs to intervene.

There is an important distinction between local-first and fully offline. Actual phone calls may require an external telephony provider, and hosted AI services may process conversation data outside the machine. Loopcall's architecture should make these boundaries explicit.

For me, open innovation is about having the freedom to choose how these components work together, rather than being locked into one closed ecosystem.

My First User

Sustainable Universe. We ran it on a real flagged order -> missing house number and pincode. Sonali opened in Hindi and asked about the rosemary conditioner. They are willing to pay after the next step which is putting the same conversation on a real phone line.

Prize Categories

Side Hustle - Challenge week - 2.
Best Use of Gemma — Gemma 4 E4B is the local brain for intent and off-script replies.

What's Next?

I want to make Loopcall more reliable beyond the demo by improving retry handling, preventing duplicate calls, strengthening response analysis, and routing unresolved cases to human operators.

The broader vision is to make practical AI automation accessible to small businesses, starting with one everyday problem: incomplete orders.

Loopcall brings the conversation and the operational outcome together. Sonali makes the call. Loopcall closes the loop.

Made by - Amrita :)

Zena - Focus and Productivity Extension for Youtube

Amrita Shrivastava — Wed, 14 Oct 2026 12:06:36 +0000

You go on Youtube, with a fixed mindset, to study something or watch a tutorial, an hour later, you find yourself half way deep into a comedy special.

This happens to the best of us, you're not alone.

But to eliminate the problem, which is, at a platform like youtube, which has all the good and the bad things, depending on your choices and type of content you consume here, you can end up with a six figure job despite no formal education or rotting your already fried brain cells.

Zena, a focus/productivity extension tool is built to tackle this very problem.

Most focus extensions hide everything or nothing. Zena gives every distraction its own switch homepage recommendations or feed, the sidebar, shorts, comments, autoplay, so the feed you get is the one you chose.

Its open-source, clone it and upload on your chrome extension to enable and use youtube distraction free!