Freshworks
Multi-Agent Report Synthesis
From weeks of analyst work to minutes — a 7-agent pipeline for executive reporting
Agent Pipeline
7-Agent
Charts Processed
50+
Time Reduction
Weeks → Minutes
The Challenge
Enterprise analytics teams were manually extracting data from 50+ charts per report, synthesising insights across metrics, and writing executive narratives — a process taking weeks per reporting cycle. The workflow was fragile, non-repeatable, and scaled poorly as Freshworks' enterprise customer base grew. There was no system for cross-metric correlation detection or multi-dimensional root-cause analysis at report level.
My Role
Led the design of the full 7-agent pipeline architecture, from metadata extraction through executive summary generation — the current engineering reference for this next-horizon product, actively in development.
- Conducted PRD gap analysis against the existing single-chart summariser
- Defined gold-standard output quality benchmarks
- Designed the prompt editor UI and cost/token tracking dashboard
- Spec'd the PDF export feature
The Solution
A 7-agent pipeline, one specialised job per agent:
- Agent 1 — Metadata Extraction
- Agent 2 — Data Validation
- Agent 3 — Trend Detection
- Agent 4 — Anomaly Flagging
- Agent 5 — Benchmark Comparison
- Agent 6 — Cross-Metric Correlation & Multi-Dimensional RCA
- Agent 7 — Executive Narrative Generation
It prioritises significant metric movements across 50+ charts simultaneously and generates executive-ready narratives with root-cause attribution — reducing analyst work from weeks to minutes.
The Outcome
Target: reduce report generation from weeks to minutes for enterprise analytics teams, extending the Freddy AI Summariser's single-chart capability to full report intelligence. Positioned as the next major monetisation lever for Freshworks' analytics platform. Gold-standard benchmarks are defined; engineering build is underway.
Key Learnings
- Agent specialisation — one job per agent — dramatically improves reliability, debuggability, and the ability to evaluate each stage independently against gold-standard outputs.
- Gold-standard output benchmarks must be defined before the build begins, not after — they are the product spec for agentic systems, not an afterthought.
- Orchestration and error handling are harder problems than the individual agents; budget at least as much design time for the seams as for the agents themselves.
Technologies & Methods
- CrewAI
- LangGraph
- OpenAI
- AgentOps
- Python
- FastAPI
- React