Backlog

problem

The Problem

Feature requests were drowning product teams

Product managers spend 5+ hours every week manually reading, classifying, and prioritizing incoming feature requests. The process is slow, inconsistent, and biased toward whoever shouts the loudest.

• 5+ hours/week lost to manual triage
• No consistent prioritization framework applied
• High-impact requests buried under noise
• Sprint planning delayed by unrefined backlogs

research

The Research

Understanding how PMs actually triage

I studied how PMs at different companies handle incoming requests — from Slack messages to formal Jira tickets. The pattern was clear: most teams lacked a systematic way to classify urgency and theme.

• Analyzed triage workflows across 10+ PM workflows
• Identified that classification by theme + urgency was the core bottleneck
• Found that clustering similar requests could eliminate 60% of duplicate work

decisions

Key Decisions

Why RAG + K-means++ over simpler approaches

I considered three approaches: rule-based classification, pure LLM generation, and RAG with clustering. Rule-based was too rigid. Pure LLM hallucinated priority scores. RAG grounded in real backlog data + K-means++ for semantic clustering gave the best results.

• Chose RAG pipeline to ground AI in real product context
• K-means++ for clustering: 10 adjustable semantic clusters
• Auto-tagging by theme and urgency instead of manual labels
• Designed for PM review — AI suggests, human decides

solution

The Solution

An AI-driven backlog refinement system

Built a SaaS tool that ingests feature requests, auto-classifies them by theme and urgency using a RAG pipeline, clusters similar requests using K-means++, and surfaces the highest-impact items for sprint planning.

• RAG pipeline for contextual classification
• K-means++ clustering across 10 adjustable semantic groups
• Priority scoring based on theme, urgency, and request frequency
• Dashboard showing clustered backlog with one-click sprint import

impact

The Impact

Deployed across 48 production instances

Backlog was deployed across 48 production instances. PMs reported faster sprint planning and better alignment between user needs and shipped features.

reflections

Reflections

What I learned and what I'd do differently

This project taught me that AI tools for PMs need to augment judgment, not replace it. The biggest insight: PMs don't want AI to decide for them — they want AI to reduce the noise so they can make better decisions faster.

• AI should suggest, humans should decide
• Clustering similar requests was more valuable than scoring them
• The 'adjustable clusters' feature was the most requested — PMs want control
• Next iteration: integrate directly with Jira/Linear APIs