AI Systems2024 – Ongoing

The Lego System

I helped build AI workflows that turn customer feedback and research data into categories, summaries, deeper analysis, and agent-supported answers.

Teams had useful feedback, but it was hard to read at scale. My work was to make that feedback easier to organise, understand, and reuse.
How To Read This

The visual shows how feedback moves through the system

Read it from left to right. Feedback comes in from different sources, the middle layers organise and reuse it, and the right side shows how people access the output through a portal or an agent.

  1. 01

    Collect customer feedback and related research information.

  2. 02

    Turn scattered comments into categories, summaries, and repeatable analysis workflows.

  3. 03

    Make the results usable through an insight portal and an agent.

Input

Feedback + research information

Processing

What the system does

01

Raw feedback

Customer comments, survey filters, research data, and product signals arrive from different places.

02

Shared structure

A taxonomy turns scattered feedback into themes that teams can compare and size.

03

Reusable workflows

Common analysis requests become repeatable workflows instead of one-off manual work.

04

Memory + agent

Useful outputs are saved so an agent can reuse past analysis when answering new questions.

Access

How teams use it

Portal

Explore, compare, understand

Agent

Ask, connect, reuse analysis

Architecture

What I helped build

Categorisation Pipeline

I built a way to group new feedback into a shared taxonomy so teams could see what topics were coming up.

Summaries

I added summaries so people did not have to read every comment inside a topic before understanding the main pain points.

Pain-Point Analysis

I turned repeated deep-dive requests into workflows for finding patterns, root causes, and evidence.

Memory Layer

I worked on saving useful outputs so later questions could reuse earlier analysis instead of starting from zero.

Portal And Agent Interfaces

Teams could use the work through a normal interface, and later through an agent that could gather relevant information.

Ship, Learn, Improve

Because production AI patterns were still new, we shipped useful versions, watched how people used them, and improved.

Evolution

Each step solved a real team problem

01Categorisation pipeline

Open feedback became easier to search, size, and compare through a shared taxonomy.

02Summarisation feature

Categories showed what topics existed, but teams still needed to understand what people were actually saying.

03Analysis workflows

After several deep dives, I saw the same steps repeating and turned them into reusable workflows.

04Agent with memory

The next step was an agent that could gather relevant information and help with analysis, not just answer one isolated question.

Scale
Cross-teamProduction adoption
4Production systems shipped
ProductExperiments informed
Reflection

I started close to research operations: fixing data workflows, helping with survey analysis, and learning how teams used feedback to make product decisions. The gap I kept seeing was simple: teams were collecting useful feedback, but much of it was still hard to read, compare, or reuse.

Because I had an NLP background, I volunteered to turn that messy feedback into something more structured. I worked with researchers to shape a taxonomy that could work across different product areas, then adapted the approach as newer AI capabilities became available.

What changed my thinking was seeing similar analysis needs appear again and again. Instead of treating each request as a one-off project, I started turning the repeated shape of the work into reusable workflows. That is how the work grew from one categorisation capability into a broader system for helping teams reason faster.

The next frontier is AI systems that remember enough to keep getting better.