Inspiration
When critical production incidents hit, on-call Site Reliability Engineers (SREs) face an overwhelming wall of context switching: digging through thousands of raw container logs, querying PagerDuty alert timelines, checking recent Jira tickets, and hunting for recent Git commit diffs.
We wanted to build an autonomous agent that acts as a first-responder Incident Commander. Instead of making engineers piece together text logs, GenBlaze Incident Sentinel automatically correlates multi-source telemetry, generates visual architecture failure/remediation topologies, synthesises a 20-second audio briefing, and commits every output to immutable Backblaze B2 storage with tamper-evident WORM provenance.
What it does
GenBlaze Incident Sentinel transforms messy telemetry into an actionable, multi-modal incident dossier in under a few seconds: Autonomous Context Discovery: Automatically fetches recent Git commits from GitHub repositories, Jira incident keys, and PagerDuty timelines via Model Context Protocol (MCP) SSE tools.
Root Cause Analysis (RCA): Leverages Gemini 3.5 Flash to analyse correlated logs, diffs, and alerts to pinpoint exact code regressions and configuration blunders.
Dynamic Visual Topology Generation: Architected with Gemini 2.5 Flash Image through a custom GenBlaze Pipeline, generating side-by-side dark-mode architecture diagrams showing system failures in glowing red and remediation paths in green.
Neural Audio Incident Briefing: Synthesises a crisp, 2-sentence executive audio summary for on-call engineers using Edge Neural TTS.
Immutable B2 Archival & Provenance: Archives all assets (PNG diagrams, MP3 audio briefings, and JSON manifests) directly to Backblaze B2 using genblaze-core and genblaze-s3, ensuring WORM compliance and SHA-256 tamper-evident provenance tracking.
How we built it
[ Unstructured Logs & Repos ]
│
▼
[ Model Context Protocol (MCP) SSE ] ──► (GitHub / Jira / PagerDuty Data)
│
▼
[ Root Cause Agent (Gemini 3.5 Flash) ]
│
┌───────┴───────────────┐
▼ ▼
[ Visual Architect Agent ] [ Audio Briefing Agent ]
(Gemini 2.5 Flash Image) (Edge Neural TTS Engine)
│ │
▼ ▼
[ Native GeminiImageProvider ] [ Backblaze B2 Object Storage ]
(genblaze-core 0.3.8 Native SDK) (PNG + MP3 + WORM Manifests)
│ ▲
└───────────────────────┘
│
▼
[ Streamlit UI on Cloud Run ] ◄────────── [ Base64 + Data URIs ]
Backend Engine: Python FastAPI microservice hosted on Google Cloud Run with NDJSON real-time progress streaming.
Agentic Integrations: Model Context Protocol (MCP) SSE Client connecting to external SRE tools.
GenBlaze SDK Integration: Upgraded to genblaze 0.4.5 / genblaze-core 0.3.8, leveraging the official GeminiImageProvider (google-gemini-image) directly in genblaze_core.Pipeline, ObjectStorageSink, and genblaze_s3.S3StorageBackend.
Multi-Modal Output: Edge Neural TTS for audio generation, Pillow for real-time asset optimization.
Frontend UI: Interactive Streamlit dashboard hosted on Google Cloud Run presenting side-by-side executive analyses, high-resolution architecture diagrams, native audio players, and Backblaze B2 S3 metadata URIs.
Challenges we ran into
Building an end-to-end multi-agent pipeline that streams high-resolution visual assets and audio to Cloud Run posed several complex engineering challenges. Here is how we discovered and resolved them:
1: Early Model Integration & Upgrading to GenBlaze Native Discovery: Early in development, pipeline orchestration required custom model adapters for Gemini image generation. Mid-hackathon, Backblaze shipped genblaze-core 0.3.8 containing the official native GeminiImageProvider (google-gemini-image).
Resolution: We immediately refactored our pipeline to adopt the native google-gemini-image provider. The seamless upgrade validated GenBlaze's zero-breaking-change architecture and allowed us to route all image generation through Backblaze's official native provider.
2: Unofficial TTS Rate Limits (HTTP 429) Discovery: Initially, we used gTTS for audio generation. When deployed to Cloud Run, shared egress IP ranges triggered Google Translate's public web rate-limiter, raising HTTP 429 exceptions that threatened to crash the streaming API endpoint.
Resolution: We refactored the audio agent to use edge-tts (Microsoft Neural Voices), saving directly to a temporary file buffer, backed by a non-blocking try...except fallback chain.
3: Streamlit Infinite Refresh Loop via WebSocket Frame Overflow Discovery: Cloud Run logs revealed frontend WebSocket frames (/_stcore/stream) connecting and immediately dropping every 2.3 seconds. Gemini 2.5 Flash Image returned raw 1.1 MB PNG diagrams. Every time Streamlit attempted to sync this 1.1 MB state over the WebSocket, it exceeded the frame threshold, dropped the socket, and caused Streamlit to infinitely reload the page.
Resolution: We introduced an automated Pillow image optimization step in main.py that thumbnail-resizes and JPEG-compresses the image down to ~120 KB (a 90% payload reduction). Combined with disabling Streamlit CORS/XSRF checks and enabling Cloud Run session affinity, the UI stabilized completely.
Accomplishments that we're proud of
Successfully battle-tested Backblaze's brand-new native GeminiImageProvider (google-gemini-image) within hours of its release.
Reduction in resolution time: Took a workflow that usually takes human SREs 30–45 minutes and automated it into a single button click.
True Multi-Modal Delivery: Delivered a complete tri-part dossier: Text (Executive RCA) + Image (Architecture Failure Topology) + Audio (Neural Briefing).
What we learned
Asset Optimization in Agent Pipelines: High-resolution AI-generated media must be dynamically compressed and adapted for real-time streaming interfaces to avoid WebSocket bottlenecks.
Autonomous Tooling with MCP: Model Context Protocol (MCP) enables LLMs to cleanly query external SaaS APIs (GitHub, Jira, PagerDuty) without hardcoding brittle API integrations into the core agent loop.
What's next for GenBlaze Incident Sentinel
Autonomous Remediation PRs: Expanding the agent to automatically draft and submit a GitHub Pull Request reversing the problematic code commit.
ChatOps Integration (Slack / Microsoft Teams): Pushing audio briefings and visual topology diagrams directly into active incident war-room channels.
Log in or sign up for Devpost to join the conversation.