Portfolio project

Voice of Customer

AI tags feedback, code flags spikes

  • 14xJump in packaging complaints, from 1 mention in July to 14 in August
  • 120Feedback items from 3 sources put into one format
  • 111 of 120Items that fit the 8 preset themes

Portfolio project, realistic test data. See all results

A direct-to-consumer (DTC) brand’s customer feedback is scattered across product reviews, post-purchase surveys and support tickets, each in its own format and with its own wording. Reading through it by hand is slow and always behind: by the time someone notices a pattern across dozens of reviews, the underlying issue (a packaging defect, a shipping problem) has often already affected weeks of orders.

Stack
n8n · Anthropic Claude API · Airtable API · Gmail · Airtable Interface Designer · Slack
Scope
Solo build: design, implementation and debugging across four pipeline stages: intake, AI tagging, trend aggregation and automated reporting
Context
Portfolio project for a fictional business (Wilder & Sage, a DTC skincare brand)

Voice of Customer

What I Built

A four-stage process turns scattered feedback into a ranked, actionable monthly digest. Every item, whichever of the three sources it came from, moves through the same four stages:

Fixed rulesStep that uses AI Hands off to the next stepReads or writes shared data
Stage 1: IntakeProduct reviews, post-purchase surveys and support tickets are put into one Raw Feedback table, with the product name added to tickets.
Stage 2: AI taggingThe AI tags each item with sentiment, severity, a supporting quote and one of eight set themes.
Stage 3: Monthly trendsEach theme’s monthly count is compared with its own history to find unusual jumps.
Stage 4: Monthly digestTop pain points, top praise, emerging themes and spikes are assembled. The AI drafts one recommended action per pain point, and the digest is emailed on the 1st.
Review queueFeedback that fits none of the eight themes is flagged with a proposed new theme name and waits here for review.
Spike alertA theme over its alert level triggers a Slack alert as soon as that month’s count runs, with real supporting quotes attached.
AirtableHolds the raw, tagged and trend tables, so every ranked theme, spike and quote links back to the original feedback.
Error handlingEvery outside step retries up to 3 times before it is allowed to fail. Any failure posts the failed step, the error and a link to the run to Slack.
Fig. 1 · The four-stage process, from raw feedback to a ranked monthly digest.

Here’s a walkthrough of the system in action:

  1. Puts three sources into one format

    Reviews, post-purchase survey responses and support tickets, each exported in a different layout, are mapped into a single Raw Feedback table.

  2. Sorts feedback into eight set themes, with room for new ones

    Each item is tagged for sentiment (positive or negative), severity and one of eight preset themes. When something doesn’t fit well, the AI flags it as a new theme and proposes a name for it instead of forcing a bad fit.

  3. Counts each theme every month and flags unusual jumps

    Each theme’s monthly count is compared with a baseline built from its own history. Once there is enough history, the alert level is its average plus two standard deviations (a statistical measure of how much it normally varies), with a minimum count so a theme can’t trigger an alert on tiny numbers.

  4. Builds and emails a monthly digest automatically

    Top pain points with AI-drafted recommended actions, top praise themes, emerging themes and spike alerts, assembled into one email and sent automatically on the first of the month.

Voice of Customer

Engineering Highlights

Select a heading to read the detail. Select it again to close it.

Rather than forcing every piece of feedback into one of eight preset categories, the AI is told to flag when nothing fits and propose its own category name. That turns a silent mistake (feedback filed under the wrong theme) into a visible queue someone can review.

The eight feedback themes with example phrases: Shipping and Delivery; Packaging; Product Efficacy and Quality; Scent and Texture; Pricing and Value; Customer Service; Ordering and Website Experience; Ingredients and Transparency.
Fig. 2 · The eight themes every feedback item is sorted into.

Each theme is compared against a baseline rebuilt from its own history every month. Once a theme has at least three earlier months, the alert level is its average plus two standard deviations (a measure of how much it normally varies). With less history it falls back to a simple rule: alert at 1.5 times the usual level. A minimum count of five means a theme can’t trigger an alert on tiny numbers.

The system flagged a genuine 14-fold jump in Packaging complaints (1 mention in July to 14 in August) automatically, in the same run it happened in. Only two earlier months existed, so it used the simple 1.5-times rule, and it attached supporting customer quotes. No one had to review the data by hand to notice it.

Voice of Customer

Results

This is a portfolio project on a fictional business: the results below are illustrative of how the system behaves on realistic test data, not measurements from live usage.

What each claim rests on

Each row matches a claim from What I Built or Engineering Highlights with the test behind it. Tested on 120 feedback items from three sources. Rows run from the strongest evidence to the weakest.

  • A fixed set of themes, plus room for new ones

    Of 120 items, 111 fit the eight themes and 9 were flagged as new themes and sent for review.

    Tested on 120 items
  • The spike it caught

    Packaging went from 1 mention in July to 14 in August and was flagged in the run that included August, with supporting customer quotes attached.

    Shown in one test
  • A fully automated monthly digest

    A sample digest for September 2026 is shown below: a spike alert, a top pain point with a recommended action and the top praise themes.

    Shown in one test
  • Spike alerts that adapt to each theme’s normal ups and downs

    The Packaging spike was caught by the simple 1.5-times rule, because only two earlier months existed. The adaptive baseline (average plus two standard deviations) and the minimum count of five were not exercised by this example.

    Designed, not tested

Limits: One spike in 120 items shows detection working once, not how often it fires or misses.

The numbers

1
Jun
1
Jul
14
Aug
Packaging complaints: a 14x jump caught automatically, the same period it occurred
Volume:
120
feedback items from 3 different sources, put into one format
Fit with the themes:
111 of 120
items fit the 8 preset themes
New themes:
9 of 120
items flagged as new themes and sent for review

A fully automated monthly digest arrives by email with ranked pain points, AI-drafted recommended actions, top praise themes and emerging issues, with no manual compiling. Every ranked theme, every flagged spike and every quote is one click from the original feedback it came from.

A sample monthly digest email for September 2026 with three sections: spikes this period (Customer Service, 3 mentions against an average of 1); a top pain point with a customer quote and a recommended action; the top praise themes.
Fig. 3 · A September sample digest, as it arrives in the inbox. It was generated before the minimum count of five was added, so it flags a spike of 3 mentions.

How this helps your business

What you get from a system like this.

  • Problems spotted sooner

    Spikes are caught the same period they occur, not weeks later when someone happens to notice a pattern across dozens of individual reviews.

  • Nothing forced into the wrong bucket

    Items that don’t fit the eight preset themes are flagged and given a proposed name instead of being filed under the wrong one.

  • One consistent format

    Reviews, surveys and tickets, each exported in a different layout, all end up in one table that can be compared side by side.

  • Fewer false alarms

    Alert levels based on each theme’s own history, plus a minimum count of five, are designed to keep small random swings from triggering false alarms.

  • Ready to act on

    The monthly digest comes with AI-drafted recommended actions attached to each ranked pain point, not just raw counts.

  • No manual compiling

    The digest is assembled and emailed automatically on the first of the month.

  • Full traceability

    Every ranked theme, every flagged spike and every quote is one click from the original feedback record it came from.

  • Early warning on emerging issues

    New themes appear in a queue for review instead of waiting for someone to spot a new pattern by hand.

Let’s find what your process is letting slip.

Book a free 30-minute call. We’ll look at one process that’s costing you time, leads or money and where it breaks down today.

Book a call