PatGPT
An innovative approach to design portfolio presentation through AI design thinking and the use of LLM tools.
Role: Product Designer
Tools: Cursor, Gemini, Supabase
Phase: Shipped August 2025
Traditional portfolio case studies are static and force users to hunt for information. I saw an opportunity to redesign the portfolio experience itself: using AI to create a conversational interface where users could ask questions about my process, projects, and design decisions interactively.
Understanding the Problem
My portfolio was getting views but not converting to interviews. People were landing on my case studies but not engaging deeply enough to start conversations despite displaying solid design work.
I concluded the issue was that my current portfolio format made hiring managers work too hard to find the specific signals they care about: my process, my rationale, and how I think through problems in a digestible manner. Ultimately, they were leaving before finding what they needed.
Building a Personalized Portfolio Assistant
I designed and built PatGPT using Cursor and Gemini, choosing to iterate in code rather than traditional Figma workflows. The key technical decision was a RAG (Retrieval-Augmented Generation) database trained exclusively on my portfolio, projects, and professional experience.
Unlike ChatGPT or other AI chatbots, PatGPT can't pull random information from the internet, it only references my actual work, ensuring users get accurate answers grounded in real documentation.
Design Considerations
Locally Stored Chats
To align with user mental models of current AI experiences, I wanted to provide users with their chat history via a side panel. I deliberated on how to best implement this feature due to the following factors: costs, technical limitations, and data privacy concerns. Ultimately, I opted to have the chats stored locally on the user’s browser to save on costs and eliminate the need to manage user data.
Prompt Suggestions
An empty input bar could potentially be overwhelming for users, so I needed to add visual affordances to quickly start conversations. I decided to help users out a bit by providing them with prompt suggestions on the homepage via chips. Once clicked, these would start a new chat with a preloaded prompt to learn more about me as a designer.
Personalized Context
While traditional LLM chatbots are extremely knowledgable, they are trained on data that would serve little purpose for users browsing my portfolio. With that constraint in mind, I utilized a RAG database to reference documents exclusively containing my portfolio projects, design decisions, and professional experience. This was to limit the amount of hallucinations and to keep the AI focused on a specific knowledge base.
Early Feedback and Iteration
I launched PatGPT v1 to a select group of users designers and engineers working on AI products. Their feedback immediately surfaced a gap I'd missed: users wanted to see visual work samples, not just text descriptions.
I got to work with my AI technical partners and implemented these changes so the RAG database could surface images of my design work within the chat feed.
Outcomes and Learnings
Product Thinking in Practice
I identified a real problem (portfolio bounce rates), designed a solution (conversational portfolio assistant), shipped it, gathered feedback, and iterated (adding visual work samples). This end-to-end process demonstrated my ability to think like a product owner by identifying opportunities and validating solutions.
Building to Understand AI
Designing in code with Cursor and Gemini gave me hands-on experience with AI/ML constraints such as context windows, retrieval accuracy, response latency. This technical understanding makes me a more effective product designer for AI features as I can design realistic solutions that account for system limitations rather than proposing concepts that can't be built.
AI-Augmented Workflow
Using Cursor and Gemini to design and build simultaneously represented a shift from traditional workflows. Rather than spending weeks in Figma before handoff, I could test UX hypotheses in real-time with functional prototypes. This experience taught me how AI tools can compress design-to-validation cycles, a capability increasingly valuable for fast-moving product teams.
Shipping Beats Perfection
I shipped PatGPT V1 quickly to gather real feedback rather than perfecting it in isolation. Users immediately revealed a gap as they wanted visual work samples, not just text. This validated my approach: ship something functional, learn from actual usage, iterate rapidly.