AI Feedback Intelligence System
A design exercise in turning unstructured feedback into structured product signal
Problem Statement
SaaS teams receive feedback across many channels: support tickets, surveys, Slack, email, but most of it stays unstructured. Without a system to process it, teams fall back on manual tagging and gut feel, which makes it hard to spot sentiment trends or prioritize the right problems. I wanted to explore how a lightweight AI pipeline could close that gap.
What I Designed
I designed an automated pipeline that ingests raw feedback, classifies it using an LLM, and routes structured insights into a dashboard without manual sorting. The goal was a real-time decision layer product teams could check instead of digging through scattered feedback threads.
The system:
- Ingests feedback from a central data store rather than scattered inboxes.
- Uses an LLM to score sentiment and categorize issues by product area.
- Generates a one-line executive summary per entry, designed for fast PM scanning.
- Syncs structured output to a lightweight database for dashboard visualization.
How It Works (System Design)
I architected this as a modular, asynchronous pipeline so each stage could fail or scale independently:
- Data Ingestion: Feedback is captured in Airtable, used as the central warehouse.
- Logic Orchestration: A Make.com scenario triggers only on unprocessed ("Blank Status") records, by design, to avoid reprocessing and unnecessary AI calls.
- AI Intelligence Layer: DeepSeek-V3 handles three tasks per entry: sentiment scoring, categorization by product domain, and abstractive one-sentence summarization.
- Data Bridging: Processed insights sync to Google Sheets as an accessible secondary store.
- Visualization: Looker Studio reads the clean data to render sentiment trends and heatmaps.
Key Design Decisions
- Chose DeepSeek over GPT-4 for the AI layer, prioritizing token cost efficiency over marginal quality gains, a deliberate cost-performance tradeoff for a high-volume use case.
- Used a status-filter trigger instead of polling on a timer, so the system only processes new records and avoids redundant API spend.
- Separated ingestion (Airtable) from reporting (Sheets) to keep the raw data layer and the dashboard-facing layer independent, making each easier to debug or swap out.
- Designed for 10,000+ monthly entries as a target scale assumption, not a measured load test, informing schema and trigger choices.
What This Demonstrates
- Translating an ambiguous operational problem ("feedback is scattered") into a concrete system architecture.
- Making explicit tradeoffs between AI cost, latency, and reliability.
- Designing for a PM's actual workflow: fast scanning and drill-down rather than just exposing raw data.