Inspiration

Have you ever been screwed over by a company? We have. According to the National Customer Rage Survey, 77% of consumers experienced a problem last year—the highest ever recorded. Yet, 96% of unhappy customers never complain. They suffer in silence.

Why? Because complaining is exhausting by design. You wait on hold for 45 minutes, send emails into the void, or face 47-field government forms. Companies spend millions making it easy to buy, and nearly impossible to complain. This creates a massive class inequality: if you're rich, you get a lawyer; if you're not, you get nothing. We built OneStop to give the "silent majority" their power back.

What it does

OneStop is not a chatbot. It is a universal execution engine that acts as your relentless, automated advocate.

You simply speak to OneStop (via our voice-first interface) and explain your problem—like an airline denying a refund due to a volcano disruption. OneStop instantly creates a complaint thread and executes a five-step escalation ladder:

  1. Intake & Strategy: An empathetic AI agent (Alex) validates your frustration and gathers necessary details (like a PNR).
  2. Email Escalation: Drafts and sends a legally sound, policy-aware email to the company.
  3. Phone Escalation: Initiates a real, outbound phone call to the company's customer service, negotiating on your behalf in real-time.
  4. Social Pressure: Drafts a high-pressure public post for X (Twitter) and tags the relevant corporate accounts.
  5. Regulatory Filing (CASE): Our proprietary "Form Genome Engine" autonomously navigates to government/consumer protection websites and fills out complex dispute forms using your case data.

How we built it

We engineered OneStop using a Swappable Provider Architecture to maximize efficiency, latency, and cost-effectiveness across a complex tech stack:

  • The Brain (Reasoning): We use OpenAI (GPT-4o) as the core state machine to handle deep legal reasoning and strategy. However, we implemented OpenRouter as a dynamic fallback layer to ensure 100% uptime during live calls, and we route high-volume classification tasks to Qwen to optimize token usage.
  • Voice & Communication: We integrated ElevenLabs for ultra-realistic Text-to-Speech, giving our agent "Alex" the perfect blend of empathy and firmness. To achieve sub-300ms latency for live phone calls, we orchestrated the pipeline through Agora. We also integrated VALSEA to handle complex Southeast Asian accents during the initial user intake.
  • Discovery & Vision: When a user complains about a company, we use Exa to scrape the web and instantly discover the correct legal entity, support emails, and regulatory bodies. We utilize JigsawStack's vOCR to parse evidence (like screenshots of denied refunds) uploaded by the user.
  • The Action Layer: Our "Form Genome Engine" uses Playwright to autonomously navigate and fill out complex web forms. To prevent government or corporate portals from blocking our bot, we route traffic through Bright Data's residential proxy network.
  • Infrastructure: The entire monolithic FastAPI backend, SQLite complaint tracker, and React/TypeScript frontend are hosted on AWS.

Challenges we ran into

Building an AI that takes action in the real world is significantly harder than building an AI that just generates text.

  • The Form Genome Problem: Government forms are notoriously brittle and change constantly. We had to build a dynamic extractor that understands the "DNA" of a form rather than relying on hardcoded CSS selectors.
  • Real-time Voice Latency: Conducting a live phone call with an airline requires sub-500ms latency. Orchestrating the STT -> LLM -> TTS pipeline required intense optimization and the use of Agora's real-time network.
  • State Management: Keeping track of a multi-channel escalation (email, phone, social, web forms) within a single unified "Thread" required building a custom durable session store.

Accomplishments that we're proud of

  • The Form Genome Engine: We successfully built an infrastructure layer that allows an AI to autonomously navigate and submit complex, multi-page regulatory forms (like CASE filings).
  • Live Phone Calls: Watching our AI successfully navigate a phone tree and argue with a customer service representative in real-time.
  • A Unified Design System: We created a beautiful, consumer-ready interface that feels like a premium workspace, not a hackathon prototype.

What we learned

We learned that the future of AI is not conversational; it is agentic. LLMs are powerful, but their true value is unlocked when they are given hands (Playwright/APIs) and a voice (ElevenLabs/Agora). We also learned the importance of dynamic model routing—relying on a single provider for a multi-step execution pipeline is a recipe for failure.

What's next for OneStop

While OneStop is starting as a consumer tool to fight back against bad corporate behavior, the underlying technology—the Form Genome Engine—is an operating system for human-to-institution interaction.

Next, we plan to expand our capabilities to handle tax disputes, insurance claims, landlord violations, medical bill disputes, and visa applications. Eventually, we see OneStop pivoting into a B2B enterprise solution, automating complex vendor chargebacks and compliance filings for large corporations.

Built With

Share this project:

Updates