PROJECT 02 · DISCOVERY / VOC
A privacy-first local LLM workflow that structures messy feedback, identifies recurring product signals, and turns them into evidence for prioritization or strengths-based decisions.
THE PROBLEM
Product Managers receive feedback through reviews, surveys, support tickets, and interviews. Without structure, teams can overreact to isolated complaints, miss recurring patterns, or prioritize from intuition rather than evidence.
HOW THE MVP WORKS
The pipeline runs locally through Ollama. Its four capabilities move raw feedback from ingestion and classification to one of two decision outputs: issue prioritization or Strength Analysis.
Reads raw feedback and converts each item into category, severity, roadmap relevance, core issue, and a recommended PM action.
Groups the local classifications to reveal recurring pain points, concentrated friction, product-health signals, and emerging opportunities.
Uses normalized frequency, severity, roadmap relevance, and manually assigned effort to turn issue patterns into comparable priorities.
When feedback is overwhelmingly positive, the workflow stops forcing fixes and identifies differentiators worth protecting and promoting.
THE PIVOT
The first version assumed that a Voice of Customer dataset should lead to a prioritized list of problems. Testing it on approximately 60 anonymized reviews from a real bakery exposed an important edge case: around 80% of the feedback was positive, so conventional issue prioritization produced limited actionable recommendations.
Instead of forcing backlog recommendations, the workflow was adapted to identify strengths, differentiators, and opportunities to explore.
PRODUCT DECISIONS
Ollama keeps the workflow privacy-first: no customer feedback needs to leave the machine for classification.
The pipeline aggregates recurring themes before applying the prioritization framework, reducing the influence of isolated feedback.
The Strength Analysis mode acknowledges that the right output is not always a backlog of fixes. The framework changes when the dataset tells a different story.
OUTCOME & LEARNING
The project produced structured classifications, PM insight reports, roadmap-prioritization outputs, and stakeholder-ready charts. More importantly, testing showed why an automated framework must accommodate context rather than mechanically convert every dataset into feature recommendations.
AI can structure recurring signals, but product judgment is still needed to interpret context and decide what deserves action.
See the local pipeline, sample datasets, generated outputs, setup instructions, and project documentation on GitHub.