Freddy AI Chart Summarizer

The Challenge

Business users — support leads, sales managers, ops heads — in Freshworks' enterprise CX and ITSM analytics platform were ignoring dashboards because charts surfaced data without narrative context. Interpreting complex chart types (stacked bars, heatmaps, treemaps) with multiple metrics required analyst involvement, creating a bottleneck that delayed decisions and weakened the platform's value proposition for Enterprise and AI-plan accounts.

My Role

End-to-end product ownership — one of six launches that earned 3× CPO recognition and $11M+ incremental ARR in 8 months.

• Wrote v1.0 and v2.0 summarisation guidelines
• Designed a six-stage pre-processing pipeline, kept separate from the LLM synthesis layer
• Built a working React prototype for engineering grooming sessions
• Defined prompt engineering standards across 5 chart types with benchmark-relative language
• Resolved UX conflicts between aggregate functions and summariser benchmark dropdowns
• Drove the multi-metric / multi-grouped enhancement that took chart coverage to 96%, and the Early Access rollout across ~25 enterprise accounts

The Solution

A GenAI-powered chart summariser with a deterministic six-stage pre-processing pipeline, kept separate from the LLM synthesis step:

• Metric extraction → benchmark calculation → trend detection → anomaly flagging → language normalisation → LLM synthesis

The system compares current metrics against user-defined baselines rather than quoting absolute numbers, producing executive-ready narratives that make charts legible without analyst involvement. Supports stacked bars, heatmaps, treemaps, line charts, and scatter plots.

Early Access now pairs the AI-generated Executive Summary and Key Findings with a conversational Freddy AI panel — a preview of where the product is headed: in-context follow-up questions on a chart's own data, not just a static summary.

The Outcome

In Early Access today across ~25 enterprise accounts, reaching 96% of user charts after the recent multi-metric / multi-grouped enhancement:

• 328 unique users engaged over the full Early Access window — +68% vs. an earlier snapshot (195 → 328)
• 2.51X average uses per user, holding steady even as the user base grew 68% — repeat usage, not one-time curiosity
• 29% of users go from AI Summary to Detailed Summary, down from 43% earlier — read as the user base diversifying beyond the initial power-user core, not disengagement, since repeat-use volume held

Recognised as a key Enterprise and AI-plan monetisation lever for Freshworks' analytics platform. Now extending in two directions at once: in-context follow-up questions on a single chart, and report-level synthesis across an entire report's key-metric movements (see Multi-Agent Report Synthesis).

Key Learnings

• Separating the deterministic pre-processing layer from LLM synthesis dramatically improves reliability and makes quality benchmarking tractable — you can test each stage independently.
• Benchmark-relative language ('8% below your Q3 average') drives better decisions than absolute numbers ('42 tickets') — context is the product, not the number.
• Building a prototype for engineering grooming is worth the investment; it reduces ambiguity, surfaces edge cases early, and halves the back-and-forth during sprint planning.
• A funnel metric that drops as the user base grows isn't automatically a regression — checking retention (2.51X held steady) before reacting to a shrinking Summary→Detail rate (43%→29%) was what separated 'broader adoption' from 'losing the power users.'

Technologies & Methods

• OpenAI GPT-4
• React
• TypeScript
• Prompt Engineering
• Tailwind CSS