Selldial

banner

Inspiration

Our team picked the VISA track because we always felt that the shopping experience for telecom products was extremely difficult for customers, based on their distinct cultural backgrounds, languages they speak, and how old they are. Speaking to T-Mobile and VISA representatives at DubHacks, we found a fusion of this problem with companies helping such consumers choose products that best suit them and manage a process to help deal with a wide range of consumers.

The problem was that the current workflow of one human dealing with hundreds of consumers with differing needs and backgrounds can get overwhelming. Our idea was having a specialized agent that can first and foremost address cultural and linguistic barriers of customers, pitch products based on their current needs and plans, and secure a checkout over the phone. More so, give control to the companies too, so they can manage how these agents are interacting with their consumers and have the authority to step in when required.

What it does

SellDial is an AI sales line for a carrier (our demo tenant is T-Mobile). A customer reaches it by phone call or message through Photon, and it:

recognizes the account and looks up their plan, lines, usage, billing and household
adapts its tone and language to the customer using a persona profile
recommends the right phone or plan and applies the right promos
suggests cheaper options when they exist, and says so honestly when waiting or changing nothing is the better choice
asks the account owner to approve purchases above their limit (\$1,500 by default)
completes the checkout securely with Visa, or with the card already on the customer's T-Mobile account when there is one, and sends a receipt that explains why it recommended what it did

A manager console shows every conversation, a decision timeline for each one (verified, recommended, owner approval, payment), customers, sales with paid and not-paid breakdowns and analytics, and settings where managers add promo codes and choose channels (voice, SMS, iMessage, WhatsApp). It comes in dark and light themes.

How we built it

Voice and channels: Photon connects the agent to phone calls and messaging, with ElevenLabs voice agents handling the spoken conversation, and webhooks into the agent runtime.
Agent runtime: each agent runs in its own Daytona sandbox, with an orchestrator that provisions, wakes and destroys it and a small in-sandbox daemon. Language-model calls go through Modal. Platform secrets stay on the host, and only an allow-listed environment enters a sandbox.
Recorded I/O: every message is written to MongoDB first (write-first), so each conversation has a complete audit trail.
Data: MongoDB Atlas. We migrated the synthetic T-Mobile hackathon dataset (devices, plans, add-ons, promotions, households with lines, usage, billing and support, and 54 knowledge rows) under its own org, with 50 persona customers, 8 prospects and 450 more households.
Console: a Next.js app serves a hand-built HTML, CSS and JavaScript console with its own design system, a dark magenta theme, a light theme, a keyboard-accessible report drawer and a mobile layout.
Payments: checkout is secured with Visa. If the customer already has a card on their T-Mobile account, the agent uses that card. Every purchase goes through the owner's spending limit and approval, and is written to the audit log.
Code review: we added VEYZL to our repo. On our pull requests it helped us notice redundancy that we then fixed, and earlier pull requests that needed to be removed and pointed to the newer ones.
Domain: selldial.biz

Accomplishments that we're proud of

Connecting Photon with ElevenLabs voice agents so customers can talk to SellDial over the phone
Connecting the voice agent to MongoDB Atlas so conversations can draw on customer accounts, plans, usage and available offers
Bringing voice and customer data together in one flow, so recommendations reflect the customer’s actual needs

Challenges we ran into

Connecting the voice agent to the T-Mobile dataset in MongoDB Atlas so it could use customer information during a conversation
Integrating the agent harness to process the data and give the voice agent the context it needed to make useful recommendations

What we learned

How we can connect Photon voice with ElevenLabs to integrate the WhatsApp agent. - Akram Jon
How to create proper evals for agents and ensure they are able to work in long running and long context environments - Avi Agola
A frontend's main purpose is not to look really great but also be feasible that is give as much interface with the user to make sure they know where to find what they need - Rishav