Skip to content
All work

USC, Viterbi School of Engineering

AMAR

A multi-agent RAG pipeline that turns plain-language questions into analytical reports.

Sep 2024 – Jun 2025

Framed for

A 50M-events/day multi-agent pipeline on Databricks and Airflow

Orchestrated RAG on Databricks with Airflow DAGs and an MLflow retraining lifecycle, stress-tested at 50M events/day against replayed public clickstream data.

50M
events/day sustained in load test
~5 min
end-to-end pipeline run
MLflow
automated retraining + tuning
  • Databricks
  • Airflow
  • MLflow
  • RAG
  • Multi-Agent
  • Python

Cutting analyst report turnaround from an hour to five minutes

Designed a system that lets non-technical stakeholders get custom analytical reports directly from a plain-language question, removing the analyst queue as a bottleneck across three business domains.

~1 hr → 5 min
report turnaround
3
business domains served
0
analyst hand-offs required
  • Multi-Agent Systems
  • RAG
  • Databricks
  • Airflow
  • MLflow

The problem

Non-technical stakeholders needed custom analytical reports but had to route every request through an analyst, costing roughly an hour per report and gating insight behind a person's queue.

What I built

  • 01Built a multi-agent RAG pipeline that parses a plain-language question, plans the analysis, and produces a custom report end to end.
  • 02Orchestrated the pipeline on Databricks with Airflow DAGs so runs were scheduled, observable, and reproducible.
  • 03Wired automated retraining and parameter tuning through the Databricks MLflow model lifecycle rather than manual redeploys.
  • 04Stress-tested at 50M events/day using replayed public clickstream data before trusting the throughput numbers.

Outcome

Report turnaround dropped from roughly an hour of manual analyst work to about five minutes, across financial-reporting, healthcare-diagnostics, and customer-support scenarios.

AMAR architecture: a plain-language question is planned, then handled by retrieval, analysis and report agents running on Databricks and scheduled by Airflow, with an MLflow retraining lifecycle, producing a custom report in about five minutes.REQUEST PATHQuestionplain languagePlanneragentRetrievalagentAnalysisagentReportagentCustom report~5 minORCHESTRATIONDatabrickscomputeAirflow DAGsschedulingMODEL LIFECYCLEMLflowautomated retraining + parameter tuningLOAD TEST50M events/dayreplayed clickstream
Request path in accent; orchestration and the retraining loop beneath. Load tested at 50M events/day before production traffic.