Context
Hiring great designers is one of the most important things I do as a design leader. It's also time-consuming and often inefficient for everyone involved.
Across hiring loops I noticed the same pattern:
- I was asking the same questions repeatedly.
- Candidates were asking the same questions repeatedly.
- Interviews often spent the first 15–20 minutes establishing context before deeper conversations could begin.
The original idea was pretty simple:
What if people could ask questions without me having to be there to answer them?
The Idea
If I could build an AI trained on my work, experience, and thinking, it could act as a scalable version of me in early hiring conversations.
Candidates could ask the AI chatbot about my experience, the products I've worked on, how I lead teams, what I look for in designers and how I think about product design. All of this on their own time, without having to fit my schedule.
As I started building it, the more interesting problem became figuring out how to turn 20 years of experience into something an AI could actually understand and talk about accurately.
The Product
I built a custom GPT-powered chatbot that allows people to ask questions about my professional experience.
The knowledge base includes content such as:
- Resume and career history
- Design case studies
- Leadership philosophy
- Past writing
- Product decision frameworks
- Experiences from companies including Step and Zynga
The interface is intentionally simple: ask a question, receive an answer, and continue the conversation. The more interesting work happens behind the interface.
Turning My Experience Into a Knowledge System
1. Convert My Articles Into Markdown
My experience wasn't originally written for an AI system. It was written for different purposes, at different times, and in different formats. The first task was to optimize that data for an AI to be able to read and understand. That meant converting it all to properly formatted markdown, adding tags, titles, etc.
2. Upload and 'Score' Relevant Articles
Embeddings are created when I upload a new article to the knowledge base. The system compares the user's question with article embeddings using cosine similarity to decide which articles are referenced in the chatbot's response.
3. Retrieve Relevant Articles
Semantic similarity is the primary scoring mechanism but additional scoring is applied when the user's question matches:
- Article titles
- Article tags
- Individual keywords within titles, tags, and content
The system then sorts the articles by their combined score and selects the five highest-scoring results.
4. Add Articles to the Prompt
The full content of the five selected articles is added to a RETRIEVED KNOWLEDGE section in the prompt.
Each article includes its title and full text. The system doesn't summarize the articles before adding them to the prompt.
This approach is relatively simple, but it creates a connection between a user's question and the parts of my experience that may be relevant to answering it.
The Prompt Layer Became the Product I Was Designing
The model doesn't automatically know how to represent me accurately. I needed to define what the chatbot was, how it should speak, and what information it was allowed to use. This evolved over time as I expanded testing beyond myself.
The system prompt instructs the chatbot to:
- Respond as an AI version of me
- Use first-person language
- Use only information in the retrieved knowledge
- Avoid making up or inferring information
- Say when the knowledge base doesn't contain a specific answer
The prompt contains the retrieved articles and a limited amount of recent conversation history as well.
The prompt isn't just instructions for the model. Getting it right is critical to creating a good user experience.
The prompt shapes the chatbot's voice, boundaries, and how the AI will respond to questions.
What I'd Explore Next
These improvements would make the experience more accurate, engaging and give the person asking the questions to go deeper.
- Show and link to the case studies and articles used to generate each response
- Test retrieval against a broader set of realistic questions
- Improve how the system communicates uncertainty
- Explore streamed responses
- Add richer formatting to improve readability
- Continue expanding and maintaining the knowledge base
Why This Matters
This project started with hiring, but it became an exploration of a bigger idea:
How could I turn a person's experience into a product?
A portfolio explains what I've done. A conversational system creates another way for people to explore how I think, what I value, and how I approach design.
The challenge isn't simply connecting a chatbot to a collection of documents. It's figuring out how to turn years of experience into a system that is actually useful and true to who you are.