JNJ SYS logoJNJ SYS Digital TechnologiesINDEPENDENT APPLICATION SAMPLE

AI AUTOMATION / CUSTOMER-SUPPORT DISCOVERY

Find repeated needs.
Before automating replies.

A working text-clustering experiment for software and chatbot teams.
Inspect the words, adjust the density, keep uncertainty visible.

TF-IDF + DBSCANSynthetic support messagesNo LLM, tickets or customer data connected

01 / THE CORPUS

What are people asking?

Higher connects less similar messages.

A core point needs this many neighbors.

02 / DENSITY-BASED GROUPS

Topics emerge. Nothing is pre-labelled.

Cards use top mean TF-IDF terms, not generated intent names. A cluster may still contain different requests; review messages before making a routing rule.

03 / INSPECT ONE MESSAGE

Why did it land here?

Most weighted terms in this message

Pairwise evidence, nearest first

Cosine similarity measures weighted word overlap, not semantic truth. The neighborhood uses 1 - similarity ≤ chosen distance.

Explore the full pairwise similarity matrix

Supplemental matrix: rows and columns use source message numbers. Darker teal means more weighted word overlap. Click a cell to inspect that message pair. Scroll inside the matrix on narrow screens.

Take the calculation with you.

Export vectors, parameters, neighbors, source messages and cluster membership.

IMPLEMENTATION / LIMITS / PRODUCTION NEXT

An observable baseline, not an autonomous helpdesk.

Implemented here

Lowercase Latin-alphanumeric tokens, a fixed stopword list, normalized term frequency, smoothed inverse-document frequency and L2 normalization. Exact cosine distances feed DBSCAN with a deterministic input-order traversal. All processing stays in the browser.

What this cannot prove

No semantic embeddings, translation or trained intent labels. Synonyms and negation can fool word overlap. Noise is not a fault or priority score. Duplicate messages influence density; corpus changes alter IDF and clusters. Border points can depend on traversal order.

Proposed production path

Authorized, redacted ticket ingestion; multilingual embedding comparison; reviewer-labelled validation; cluster stability and drift checks; access controls. Only reviewed patterns become candidate chatbot or routing rules. No support reply is sent.

Independent sample by Pranay Bhardwaj, informed by JNJ SYS's AI automation, chatbot and software services. All supplied messages are invented. No measured customer benefit is claimed.