Customer Knowledge Graph

The problem

Customer feedback arrives as unstructured noise across support tickets, calls, reviews and surveys. Manual tagging doesn't scale and taxonomies go stale the moment the product changes. Worse, classification alone answers what customers said — not who said it, which accounts it puts at risk, or what revenue it touches.

Layer 1 — the knowledge graph

The graph is the foundation: it connects every piece of feedback to products, users, accounts and opportunities, so every insight is automatically grounded in context.

• → Feedback is automatically linked to product, user, account and opportunity metadata — full traceability from a complaint to the revenue it touches.
• → Relationships between feedback, features, customers and opportunities enable questions like “what are enterprise customers saying about onboarding?”
• → Source-specific fields — Account ARR, user plan type, opportunity stage — are retained, searchable and filterable.

The entity model

Layer 2 — the adaptive taxonomy

On top of the graph sits a living, five-tier classification structure that keeps feedback categorization accurate and aligned with how the business evolves.

• → Five tiers — L1, L2, L3, Theme and Subtheme — for analysis and reporting at any altitude.
• → Themes link to product features, so positive and negative trends appear in the same view.
• → Detects and merges overlapping topics, keeping the taxonomy clean as it grows.
• → Learns each company's data and language automatically — reducing manual tagging and taxonomy drift to zero.

All five tiers at once — imagine Spotify's feedback flowing through it

L1 → L2 → L3 narrow the topic; Themes carry sentiment; Subthemes hold the specifics. Every count updates as new feedback lands — no human retagging.

One line: the taxonomy tells you what's being said; the graph tells you who's saying it and what it's worth.

What the graph makes possible

The graph retains source-specific fields — Account ARR, user plan type, opportunity stage — and keeps them searchable and filterable. That turns vague dashboard questions into precise, answerable ones:

• → “What are enterprise customers saying about onboarding?”
• → “Which complaints sit on accounts with open six-figure renewals?”
• → “What feedback themes correlate with at-risk health scores?”

Try the idea

A miniature of the graph — hover any node to trace how feedback connects to product areas, accounts and revenue:

50+ sources

“Exports keep timing out”

Feedback24.5k

AccountAcme

SourceZendesk

TypeComplaint

User8.2k

NameAlex Doe

PlanEnterprise

Account1.4k

ARR$500,000

HealthAt risk

Opportunity320

Value$500,000

StageNegotiation

1

Collect

feedback pours in from 50+ sources

2

Connect

linked to the user, account & revenue behind it

3

Classify

auto-tagged by a taxonomy that keeps learning

Inside the taxonomy · a music app's feedback, illustrative

L1Playlists18,916

└L2Collaborative Playlists742

└L3Invites & Sharing416

five tiers narrow 24.5k records to the exact topic

Invite links expire too quickly

Add QR-code invites

“Editing together just works”

themes carry sentiment — complaint, ask, praise

My role

I worked on the taxonomy and graph end to end — the graph data model, how the semantic classification layer sits on top of it, and the continuous-learning loop that keeps the taxonomy aligned with each customer's changing product. For scope honesty: this foundation later shipped publicly as part of Enterpret 2.0 after my time there — I claim the foundational work, not the 2.0 launch.