Product Adoption & Churn Intelligence Assistant
An end-to-end agentic decision system that analyzes customer usage signals to surface feature adoption opportunities, actionable enablement playbooks, and explainable churn risk, translating raw telemetry into CSM-ready guidance.
Tags & Technologies
Key Impact & KPIs
- Under-utilized feature opportunities surfaced with clear business rationale
- Churn risk signals consolidated into explainable Low / Medium / High classifications
- Reduced manual analysis for Customer Success and Product teams
- Faster, more consistent customer enablement decisions
- Trust-first recommendations through deterministic, auditable logic
Project Overview
1. Agentic Decision Pipeline
Designed an agentic decision pipeline that unifies customer usage telemetry, feature exposure, and behavioral signals into a single assistant—enabling Customer Success teams to move from raw activity data to prioritized, action-ready recommendations.
2. Modular Agent Core Architecture
Built a modular Agent Core and tool-based architecture that orchestrates discrete capabilities (usage summarization, feature exposure analysis, recommendation ranking, churn signal detection), ensuring clarity, extensibility, and production-readiness without monolithic logic.
3. Explainable Adoption Recommendations
Implemented explainable, rules-driven adoption recommendations that identify high-impact, under-utilized features and generate a concise 1–3 step enablement playbook for the top opportunity—preserving transparency over black-box scoring.
4. Interpretable Churn-Risk Framework
Developed an interpretable churn-risk framework that consolidates multiple behavioral indicators (engagement trends, inactivity patterns, support signals) into human-readable Low / Medium / High risk labels with explicit reasoning for each classification.
5. Demo-Ready Extensible System
Delivered a demo-ready, extensible system with typed data models, in-repo mock telemetry, memory/context tracking, and dual interfaces (CLI + Streamlit), demonstrating how agentic analytics systems can be operationalized responsibly in customer- and revenue-sensitive environments.
Model Selection Rationale
- Models/LLMs used: Cohort analytics + XGBoost for adoption scoring; Flan-T5-Small to synthesize multi-metric narratives.
- Actionability & speed: XGBoost delivered stable, fast scoring for operational pipelines; Flan-T5-Small produced concise insights at low latency.
- Practicality: Cohort-first approach preserved temporal context; models focused on interpretable signals product teams can act on.