TraceTalk: AI Agent for Live System Observability

A chat interface where you ask natural language questions about your microservices and an AI agent queries OpenObserve logs in real-time to investigate and answer.

@ogbhau

github.com/Omkar-Ghongade/Push2Prodvisit ↗

TraceTalk: AI Agent for Live System Observability

process.log7 fields

toolsClaude OpusOpenObserveNext.jsFlaskAnthropic API

iterationsMany attempts

debugging agent loops, fixing SSE streaming, proxy issues, markdown rendering

time2 Hours

%_ai70

me 30

ai 70

self-reported · generously rounded

# how_it_was_made

proof:Devfolio↗

questions0

no questions yet

ask how it was made — the dead-ends are usually the good part

ask about the process…sign in

~/slopfolio··

TraceTalk lets SREs ask questions like "How is my database health?" and get detailed answers with evidence from live logs. I built the backend agent loop in Python — it receives a question, calls Claude with tool definitions (run_sql, search_logs, list_streams), executes the SQL against OpenObserve, feeds results back to Claude, and streams narration + final answers via Server-Sent Events. The AI wrote the agent loop, Flask SSE endpoint, and Next.js frontend with glassmorphism styling inspired by Kree8's Dribbble work. I directed the architecture choices: using application_logs instead of metric streams (which return 400 errors on SQL queries), increasing the tool call budget from 5 to 20 so the agent finishes investigations, and bypassing Next.js proxy to call Flask directly to fix timeout issues.

// best_dead_end

The agent kept trying to query metric-type streams (db_query_duration_ms, http_latency_p99_ms) which OpenObserve rejects with 400 errors. Fixed by updating the system prompt to explicitly tell Claude that only application_logs is queryable via SQL.

# prompting_approach

Started with functional requirements ("build an agent that queries OpenObserve"), then iteratively debugged specific failures ("getting 405 error", "tables showing as raw markdown"). When the UI needed polish, I pointed to a specific designer's Dribbble portfolio and asked to match that style.