CloudSentry AI Fleet

An autonomous, zero-trust multi-agent SRE fleet powered by Gemini 3.5 Flash and Google ADK that diagnoses, sanitizes, and remediates Google Cloud infrastructure incidents in real-time.

Comment

Inspiration

Film production is notoriously slowed down by the visual effects (VFX) pre-visualization phase. Translating a director's vision into initial CGI blockouts or procedural shader scripts typically takes hours of manual setup in software like Blender or Three.js.

In the era of Agentic AI, writing code by hand should not be the bottleneck in media production. We realized that if we could equip Google Cloud Gemini Enterprise agents with a real-time, sandboxed execution environment, we could transform natural-language directorial instructions into live, rendered VFX pre-visualization pipelines instantaneously.

What it does

FrameForge AI acts as an autonomous visual effects crew member inside a multi-agent cinema studio.

Directorial Interpretation: A Director Agent parses scene directions (e.g., "Create a futuristic neon city street with rainy reflection effects and volumetric lighting").

Dynamic Script Synthesis: Gemini Enterprise synthesizes complete, executable Python render scripts (using Three.js / PyOpenGL / Blender Python API).

Replit Sandbox Execution: FrameForge connects to Replit via a custom Model Context Protocol (MCP) server, provisions an isolated cloud environment at runtime, runs the generated script, and validates the visual output.

Live Studio Preview: Delivers a playable video preview and interactive web canvas back to the studio interface in seconds.

How we built it

Orchestration & Intelligence: Built using Google Cloud Agent Builder and Gemini Enterprise (gemini-2.5-flash) to manage tool execution, task decomposition, and code generation.

Runtime Sandbox Execution (Replit Track): Designed a dedicated FastMCP server (mcp_servers/replit_mcp/server.py) that interacts directly with Replit's environment APIs at runtime. Gemini agents dynamically invoke tools like create_vfx_sandbox and execute_scene_script to execute render pipelines live inside Replit environments.

Backend Infrastructure: Developed in Python using FastAPI for real-time web-socket communication and Pydantic for structured data validation.

Frontend Studio Interface: A responsive real-time dashboard built with Tailwind CSS and HTML5 video streaming, providing live logs of agent reasoning and sandbox execution.

Challenges we ran into

Deterministic Code Execution in LLM Pipelines: Generative code can occasionally contain import errors or missing dependencies. To solve this, we implemented an autonomous self-correction loop where error outputs captured from the Replit execution sandbox are fed back into Gemini Enterprise to automatically debug and rerun the script.

MCP Protocol Integration: Ensuring low-latency, bidirectional tool calls between the Gemini Agent Builder orchestration layer and the Replit MCP server required careful schema alignment and asynchronous client routing.

Accomplishments that we're proud of

True Runtime Integration: FrameForge does not just display code—it provisions real cloud environments and executes live scripts inside Replit at runtime, fully meeting the core requirement for the Replit Track.

Autonomous Error Recovery: Successfully demonstrated self-healing multi-agent behavior where Gemini fixes runtime Python exceptions inside the Replit sandbox without human intervention.

Seamless Pre-Visualization: Reduced the time required to generate scene CGI blockouts from hours to less than 30 seconds.

What we learned

The Power of Model Context Protocol (MCP): MCP provides a standardized bridge that allows LLMs to safely interact with external compute engines like Replit as if they were native system tools.

Sandboxed Execution Security: Leveraging Replit's infrastructure allows autonomous AI agents to write and execute arbitrary Python code safely without risking system corruption on the main host server.

What's next for CloudSentry AI Fleet

Blender & Unreal Engine API Deep Integration: Expanding the Replit sandbox runtime environments to support headless Blender rendering and Unreal Engine Python automation.

ClickHouse Scene Asset Indexing: Integrating ClickHouse as a persistent vector memory store to index past rendered VFX assets for instant retrieval across multi-episode productions.

Multi-Agent VFX Crew: Adding specialized lighting, camera motion, and texture agents that collaborate in parallel inside shared Replit workspace sandboxes.

Built With

Share this project:

Updates

posted an update —

Project Log: The Evolution of CloudSentry AI Fleet

Phase 1: Concept & Problem Identification Modern cloud infrastructure monitoring creates massive log noise during active incidents. Standard incident response pipelines frequently expose raw telemetry, API keys, and JWT tokens across internal monitoring channels. CloudSentry AI Fleet was conceived as a zero-trust, multi-agent control plane designed to intercept, sanitize, analyze, and remediate infrastructure threats in real time without leaking credentials.

Phase 2: Multi-Agent Topology & Model Armor We designed a four-agent orchestration mesh powered by Gemini 3.5 Flash and the Google Agent Development Kit (ADK):

Ingress Guard Agent: Parses incoming log streams and APM traces.

Model Armor Agent: Intercepts raw payloads using regex and guardrails to strip sensitive PII, DB passwords, and authorization headers before sending context to LLMs.

Zero-Trust Policy Agent: Leverages Gemini 3.5 Flash for rapid Root Cause Analysis (RCA) and dynamic severity scoring.

Remediation Sandbox: Auto-generates execution-ready code patches.

Terraform

Auto-Generated Remediation Patch (Terraform HCL)

resource "google_compute_firewall" "isolate_compromised_workload" { name = "deny-sql-injection-vector" network = "default"

deny { protocol = "tcp" ports = ["80", "443"] }

source_ranges = ["192.0.2.0/24"] target_tags = ["quarantine"] } Phase 3: Interactive Control Plane & Real-Time Analytics We constructed the Streamlit dashboard to provide full visibility across the incident lifecycle:

Timeline Stream: Live status dropdowns tracking agent execution steps.

Real-Time Analytics: Visualizing threat vector distribution, resolution latency, and cumulative neutralizations.

Generated Code Patch: Real-time Terraform HCL firewall rules and gcloud CLI commands.

Raw Audit JSON: Complete structured logs for compliance post-mortems.

Phase 4: Cloud Run Deployment & Benchmark Milestone

Deployed containerized microservices directly to Google Cloud Run with Google Secret Manager for key management.

Achieved an average Mean Time To Resolve (MTTR) of 1.2 seconds per telemetry incident.

Completed full submission packaging for the Google Cloud #AllThingsAgentic Hackathon!

Log in or sign up for Devpost to join the conversation.

Submission history