Downstream API Status: HEALTHY (200)
Downstream API Status: DOWN (503)

Inspiration

In high-throughput microservice architectures (such as payments, e-commerce, and logistics engines), sudden database connection drops or downstream API latency spikes usually lead to catastrophic failure chains. Traditional circuit breakers simply throw a 503 Service Unavailable error, leading to lost transactions and degraded user experience. Inspired by high-reliability transaction processors like Razorpay's Vulcan architecture, we wanted to build a resilient, agentic safety net using Gemini and Google Cloud that transparently heals backend failures without dropping state or crashing end-user flows.

What it does

Vulcan-Shield is an autonomous AI agent layer that acts as a dynamic circuit breaker and dynamic fallback environment for mission-critical microservices:

Real-Time Fault Interception: Continuously monitors serverless/microservice execution streams and intercepts 5xx infrastructure errors or network timeouts before they hit the client.
Dynamic Mock Environment Synthesis: Using Gemini 1.5, the agent analyzes the incoming payload, failed API endpoint schema, and historical system context to instantaneously synthesize a zero-state, compliant mock environment, returning a valid provisional payload to keep the user flow alive.
Autonomous Queueing & State Replay: The agent isolates mutated backend events inside a secure staging queue (Google Cloud Pub/Sub) and actively polls down-stream health nodes. Once the target microservice recovers, the agent autonomously replays queued mutations, reconciling state asynchronously.

How we built it

Intelligence Layer: Built with Gemini 1.5 via the Google GenAI SDK, leveraging Structured Outputs and Function Calling to inspect payload structures and generate accurate mock API fallbacks.
Orchestration & Workflow: Developed using Google Cloud Workflows and Python fast-api microservices to intercept service layer exceptions.
Data & Queueing: Powered by Google Cloud Pub/Sub for event-driven message queuing and transaction tracking.
Host Infrastructure: Deployed on Google Cloud Run to handle high-throughput, low-latency agentic execution.

Challenges we ran into

Sub-Second Latency Requirements: Standard LLM calls can be too slow for transaction microservices. We optimized Gemini's system instructions and used structured JSON schemas to ensure fallback synthesis completes within milliseconds.
Ensuring Idempotency on Replay: Preventing duplicate database mutations when replaying queued transactions after downstream nodes healed required strict transaction signature generation.

Accomplishments that we're proud of

Built a zero-downtime microservice safety network that gracefully handles downstream database crashes.
Successfully executed end-to-end self-healing cycles without losing a single transaction state during simulated load tests.
Designed a fully decoupled architecture adhering to Google Cloud best practices.

What we learned

How to harness Gemini's structured function calling for non-chat, high-utility background execution tasks.
Strategies for designing robust, agentic fallback loops that prevent cascading microservice outages in enterprise applications.

What's next for Vulcan-Shield

Multi-Cloud Node Health Prediction: Integrating predictive latency telemetry to reroute transactions before failures happen.
Automated Root-Cause Analysis (RCA): Enabling the agent to generate detailed post-mortem reports and GitHub pull requests to patch code bugs automatically.