ProofPilot

ProofPilot uses Google ADK & Elastic to turn API docs & customer context into functional, source-grounded React+Node demos in seconds, with built-in claim checking to avoid hallucinations.

Comment

Video Demo: https://drive.google.com/file/d/1cGMlf2d2TCeCminnSx0YJvmeubZ45OS4/view

Inspiration

Sales engineers, solution architects, and technical writers spend countless hours manually handcrafting custom API demos to show product fit. At the same time, buyers struggle to verify if API marketing promises hold up against raw documentation and their internal business constraints. We wanted to build a tool that automates this entire pipeline, generating highly tailored, functional codebases in seconds while proving with absolute certainty that every claim made is 100% grounded in technical reality. ## What it does ProofPilot is an agentic, source-grounded API demo generator. Users enter an API name, raw documentation or a public docs URL, a customer name, and a free-form customer scenario. ProofPilot then:

  1. Indexes documentation and customer context.
  2. Extracts specific capabilities, business signals, and customer pain points.
  3. Generates a tailored React frontend and Node backend proxy demo package.
  4. Validates all demo claims using a dedicated Claim Checker Agent that cross- references assertions against indexed documentation, pruning or rewriting unsupported statements to prevent hallucinations.
  5. Packages and commits the codebase directly to GitLab as a reviewable merge request with a live browser preview.

## How we built it

• Frontend: Built with Vite React and Vanilla CSS, featuring a responsive, dashboard- style UI displaying a real-time progress route map of active agents. • Backend: Express Node.js service managing the named agent workflow. • Orchestration: Built using the Google Agent Development Kit (ADK) and Gemini (Vertex AI), utilizing dynamic skill loading to register specialized capabilities on demand. • Grounding Layer: Elasticsearch (using Elastic Agent Builder tools over MCP) indexes raw docs and customer context, serving as the source-of-truth knowledge base. • Deployment: Containerized and deployed to GCP Cloud Run utilizing private service- account authorization via Google's Metadata Server.

## Challenges we ran into

• Metadata Server Hangs: When running the Cloud Run backend locally, identity token requests to metadata.google.internal would hang indefinitely. We resolved this by implementing a fast-abort AbortController timeout that immediately falls back to local mock credentials. • Simulation Synchronization: Keeping the frontend's station progress indicators aligned with the asynchronous API agent trace was challenging. We solved this by refactoring the React state transition logic so that already-completed stages retain their completed description rather than flipping back to pending. • Context Adaptation: Ensuring that the local fallback heuristics generated correct dashboards (e.g. Open-Meteo variables like temperature and coordinates) rather than defaulting to hallucinations.

## Accomplishments that we're proud of

• Architecting a multi-agent workflow where Google ADK and Elastic Agent Builder MCP tools seamlessly interact. • Creating a secure, private backend hosting setup on GCP Cloud Run.

## What we learned

• Dynamic Skill Compacting: We learned how to dynamically register tools using ADK SkillToolset objects to prevent prompt context bloat. Learnt how to use Elastic Search to make our lives easier! • Hybrid Testing: Designing mock-state engines that mimic complex LLM behaviors is critical for offline/local developer productivity.

## What's next for ProofPilot

• Live Sandbox Verification: Add a sandboxed execution environment to run npm run build and test the generated React applications automatically before committing. • Hybrid Retrieval: Integrate dense vector embeddings alongside BM25 text search inside the Elasticsearch grounding adapter for even more precise signal mapping.

Built With

Share this project:

Updates