AI ACCOUNT INTELLIGENCE
A repo-backed customer-success command center that brings account health, product usage, support context, issue tracking, CRM-style ownership, and AI-assisted account chat into one operating view using fictional account data.
← Back to ProjectsThis project explores a common customer-success problem: account context lives across usage dashboards, support systems, issue trackers, CRM records, and individual memory.
It uses fictional customers and fixture data to show how those signals can be gathered into one review surface with an AI layer that answers account questions and suggests a next step.
Portfolio focus: Customer-success product design, account-health synthesis, LLM tool use, and pragmatic boundaries between demo data and product behavior.
Usage, ARR, renewal timing, subscores, and monthly trend data.
Recent tickets and ticket titles available through chat tools.
Customer-tagged issues, open counts, latest issue, and velocity.
Snapshot-first answers with bounded tool use for deeper context.
The app is built around the decisions a CSM needs to make: where to spend time, what changed, which accounts are at risk, and what to say or do next.
The list view supports sorting, searching, health pills, and CSM filtering; the detail view puts usage, support, engineering, and contract context side by side.
The page model assumes a Next.js app with a fixture data layer, mock source-system clients, an account chat route, a next-step recommendation route, and an auth stub.
The important boundary is that UI components depend on stable account, trend, ticket, and issue interfaces rather than on any one backend implementation.
Snapshot-first LLM design. The chat sidebar embeds the account snapshot as cache-controlled context and only calls tools for details the snapshot cannot carry.
Bounded tool use. The assistant can list support tickets, customer-tagged issues, or account trends, with a max-turn cap to control cost.
Pre-computed next-step signals. Renewal proximity, utilization level, weakest subscore, issue velocity, and usage flags are calculated before the model sees them.
Demo-safe data boundaries. The project keeps fictional source data behind generic interfaces so the product behavior can be shown without exposing real systems.