OpsPilot AI

Dashboard

Inspiration

In every software company, when a production system breaks at 2 AM, an engineer has to manually check logs, metrics, and deployments — taking 60 to 90 minutes every single time. The same incidents repeat because there is no institutional memory. We built OpsPilot AI to eliminate this completely.

What it does

OpsPilot AI is a self-learning autonomous incident commander. When a production alert fires:

Recalls past incidents from Elasticsearch memory using semantic search
Analyzes live logs using ES|QL queries via Elastic Agent Builder MCP tools
Correlates CPU, memory, and latency metrics using ES|QL
Investigates recent deployments as potential root causes
Reasons over all evidence using Gemini 2.5 Flash to identify root cause
Executes remediation automatically — rollback, scale-up, or restart
Notifies the team via real Slack alert with full incident details
Creates a Jira ticket automatically with root cause and actions taken
Stores learnings back into Elasticsearch so future incidents resolve faster

What takes an engineer 60-90 minutes manually, OpsPilot completes autonomously in under 2 minutes — with zero human intervention.

How we built it

Frontend: Next.js 14 + TailwindCSS — live reasoning trace dashboard with WebSocket streaming

Backend: FastAPI (Python) — orchestration, REST API, and WebSocket server

AI Engine: Gemini 2.5 Flash with function calling — autonomous multi-step reasoning

MCP Integration: Elastic Agent Builder MCP Server — all data retrieval goes through verified MCP tools:

platform_core_search — semantic search using ELSER
platform_core_generate_esql — converts plain English to ES|QL
platform_core_execute_esql — runs queries against live Elasticsearch data

Database + Memory: Elasticsearch Serverless — stores logs, metrics, incidents, and agent learnings

Real Integrations: Slack webhooks for real-time alerts, Jira REST API for automatic ticket creation

Challenges we ran into

Elastic Cloud Serverless rejects number_of_shards and number_of_replicas in index mappings — had to remove all index settings
MCP server requires Accept: application/json header specifically — text/event-stream alone returns 406
Ensuring all decisions come from Gemini reasoning rather than hardcoded if/else logic — rewrote all agents to return raw data and let Gemini interpret it
Gemini free tier rate limits — switched to models/gemini-2.5-flash with proper model path format

Accomplishments that we're proud of

Truly autonomous — zero human input from alert to resolution
Self-learning memory — every resolved incident stored in Elasticsearch, recalled by future incidents
All retrieval through MCP — agents never query Elasticsearch directly, always through Elastic Agent Builder MCP tools
Real integrations — Slack messages and Jira tickets created automatically, not simulated
Live reasoning trace — dashboard streams every Gemini thought and tool call in real time via WebSocket
Decision transparency — every root cause and remediation choice is made by Gemini, visible on dashboard

What we learned

How to build truly autonomous agents using Gemini function calling
How Elastic Agent Builder MCP server exposes enterprise data to AI agents
How ES|QL enables AI agents to query structured data using natural language
How Elasticsearch serves as persistent agent memory, not just a log store
The difference between a chatbot that answers questions and an agent that takes actions