P
Back to Case Studies
Featured Case Study

Multi-Agent System for IT Operations Automation

Orchestrated agents for incident detection, triage, and auto-remediation at scale — with human-in-the-loop guardrails and full traceability.

Multi-AgentLangGraphKafka Healthcare
70%
Faster Incident Turnaround
60%
Processing Efficiency Uplift
10K+
Events Handled Monthly
100%
Critical Actions Traceable

Business Problem

Long-running operational processes failed in partial and unpredictable ways. A single prompt chain cannot safely manage event variance, external tool calls, validation, review, retries, and downstream updates.

Architecture Overview

A stateful multi-agent workflow that separates deterministic processing from agentic decisions, persists checkpoints, and makes every transition reviewable.

  1. 1 Event sources (alerts, tickets, metrics, logs)
  2. 2 Ingestion + event collection
  3. 3 Supervisor agent orchestration
  4. 4 Specialist agents (triage, root cause, resolution, knowledge)
  5. 5 Human review gate
  6. 6 Actions + audit (tickets, runbooks, notifications)

All components are deployed with private networking, scoped IAM roles, and encryption at rest and in transit.

Key Design Decisions

  • Workflow first, agent where useful. Deterministic nodes own policy and state transitions; agents handle bounded ambiguity with explicit tools.
  • State outside the model. Durable checkpoints make long-running jobs resumable and prevent conversation history from becoming the system of record.
  • Human review is designed in. Confidence, policy, and value thresholds route only the right cases to reviewers with the evidence needed to decide.

Technology Stack

LangGraphKafkaPostgreSQLRedisFastAPIPython

My Role

  • Designed the supervisor + specialist agent architecture.
  • Implemented durable state management and checkpointing.
  • Built typed tool interfaces and structured agent handoffs.
  • Designed the human-in-the-loop review gates and queues.
  • Instrumented per-agent cost and trace observability.

Measurable Impact

  • 70% Faster Incident Turnaround
  • 60% Processing Efficiency Uplift
  • 10K+ Events Handled Monthly
  • 100% Critical Actions Traceable

Challenges & How I Solved Them

  • Partial failures: Node-level retry and compensation policies with duplicate-execution protection around side effects.
  • Inter-agent drift: Structured, schema-validated outputs between agents replaced free-text handoffs.
  • Cost control: Per-agent token budgets and model routing halved run costs without quality loss.
  • Trust and audit: Every transition, tool call, and approval is traced and reviewable.

Lessons Learned

  • Keep agent responsibilities narrow and well-defined.
  • Prefer deterministic workflow nodes; add agents only where ambiguity needs judgment.
  • Checkpoint after side effects, not before.
  • Structured outputs between agents prevent systems from rotting.