2025
Segmentation Agent
A multi-agent system that turns a raw customer CSV into clusters, segment profiles, and a marketing playbook.
Context
Customer segmentation usually lives in a notebook a marketer can't open. I wanted to compress that whole pipeline (preprocess, cluster, validate, profile, recommend) into something a non-technical user can run on their own data and then ask follow-up questions about.
Approach
- Built a sequential pipeline on Google's Agent Development Kit. A Clustering Agent calls custom Python tools, then a Marketing Strategy Agent consumes its output.
- MiniBatch K-Means with an elbow-method selector based on geometric perpendicular distance from the inertia curve, plus Silhouette, Davies-Bouldin, and Calinski-Harabasz cross-checks.
- A separate Q&A Agent keeps conversational context, so users can ask 'how did you pick k?' or 'which segment should I prioritise as a small business?' after the run finishes.
- Streamlit front end with cached runners and DOCX/CSV report exports.
Outcome
An end-to-end run on an uploaded CSV produces three downloads (clustering report, marketing strategies, raw data with cluster labels) plus a chat thread for follow-up questions. It demonstrates agent orchestration, tool design, and grounding LLM output in real numerical analysis.
Stack
- Google ADK
- Gemini 2.5
- Python
- scikit-learn
- Streamlit