honkler

AI voice agents that argue to get you better deals on your bills.

Comment

Inspiration

Americans overpay on their bills every single month, roughly $400+ per household per year. The kicker is that 80% of people never even try to negotiate, and the ones who do spend an average of 3 hours haggling with customer service. Existing solutions like Trim and Billshark rely on human agents, which makes them slow (days to weeks) and expensive (40–60% of savings). We wanted to build something that removes the human bottleneck entirely: an AI agent that handles the full negotiation call, end to end, in minutes.

What it does

Honkler lets you upload a photo or PDF of your phone, cable, or internet bill. Our system analyzes your current plan, researches competitor offers and available loyalty discounts, then deploys an AI voice agent that calls your provider and negotiates a lower rate live. The entire pipeline from upload to savings is fully automated, with no human in the loop.

How we built it

We engineered Honkler with a three-tier distributed system spanning a React 19 + TypeScript frontend, an Express.js API backend with PostgreSQL (via Prisma ORM), and a Python FastAPI agent runtime that orchestrates the core AI pipeline. When a user uploads a bill, it kicks off a multi-stage AI research pipeline powered by Google Gemini 2.5 Flash. First, it performs multimodal document extraction on the uploaded PDF/image, then runs a live competitor market research using Gemini's Google Search grounding, and finally synthesizes a full negotiation strategy with structured JSON output via Pydantic-validated schemas. Research progress streams to the frontend in real time through Server-Sent Events (SSE), including Gemini's extended thinking tokens so users can watch the AI reason step-by-step. Once the strategy is ready, we initiate an autonomous phone call through ElevenLabs' conversational voice AI integrated with Twilio, dynamically injecting the negotiation plan into the agent's prompt at call time. Live call transcripts are broadcast to the frontend over WebSockets, giving users a real-time window into the negotiation as it happens. After the call, Gemini analyzes the full transcript against the original strategy to determine outcomes and calculate savings, which are then settled on-chain via Solana using the x402 payment protocol. The entire system is held together with end-to-end type safety (TypeScript on the frontend and API, Python type hints and Pydantic on the agent), JWT authentication, and a robust error-handling layer with custom exception hierarchies across both runtimes. File storage is handled on AWS, and the agent runtime is on Vultr.

Challenges we ran into

I'm most proud of the real-time experience. Users can watch the AI "think" during the research phase via streamed Gemini thinking tokens, then listen to a live transcript of the actual phone call over WebSockets. It makes the whole process feel transparent rather than like a black box.

Accomplishments that we're proud of

I learned that orchestrating multiple AI services in real time is a fundamentally different problem than calling an API and waiting for a response. Managing async streams from Gemini, live WebSocket transcripts from ElevenLabs, and Twilio webhook callbacks simultaneously taught me a lot about event-driven architecture and the importance of robust state management. Working with Solana's x402 payment protocol for settlement was new territory for me, and it showed me how blockchain can fit naturally into a service-based product when the use case is right.

What we learned

I learned that the hardest part of building an AI-powered product isn't any single integration, it's orchestrating all of them together. Getting Gemini, ElevenLabs, Twilio, and Solana to work in harmony, each with their own async patterns, webhooks, and failure modes, forced me to think deeply about system design in a way I hadn't before.

What's next for honkler

Expanding beyond telecom to insurance, medical bills, and subscription services. Building a recurring monitoring system that automatically renegotiates when your promotional rate expires. And scaling the voice agent to handle multi-party calls and escalation paths.

Built With

Share this project:

Updates