I added a ChatGPT-style assistant to my portfolio so visitors could ask about me. $5 to start, markdown for context, Neon to track usage. A way to stand out and keep building while job searching.
I found out I was getting laid off from Axure six days ago. I started this website the day I found out. I had always wanted to build my own website. Now I had a reason and a deadline: stand out while looking for the next role. I had a lot of ideas. One felt both relevant and a little bold: what if my site had its own way to "talk to me"? Not a contact form. A chatbot that knew about me and could point people to my work. Would it be super useful? Probably not. Would it be novel and would I learn a ton? Yes. So I built the chat interface today.
I wanted the chatbot to help get me a job. Really. I wanted it to refer visitors to the right case studies and resources so that the time I put into those pieces actually got seen. If the bot could redirect people to the work I care about, it would have done its job. I also wanted to keep building on top of the job search. Creating and shipping something would keep me sane.
Backend: Nuxt server route that streams from Claude (Sonnet 4.5). System prompt tells the model it's my assistant, to keep answers short, to stick to the context file, and to suggest case studies. Context is one markdown file (about me, resume, tools, case study summaries). I kept it simple so I could ship without adding a whole new stack.
Homepage hero with a textarea and send button that routes to /chat?q=... . Chat page: back link, message list (user bubbles and plain assistant responses), and an input with send button. Styled to match the rest of the site. Enter to send, Shift+Enter for newline.
In-memory rate limit per IP (e.g. 20 requests per minute) so one user can't blow up the Anthropic bill. Returns 429 with a message when over limit.
I wanted to know if anyone ever used it. Optional Neon Postgres: when NUXT_DATABASE_URL is set, each user message (and optionally the assistant response) is written to a chat_submissions table. No auth, just a log so I can see what people ask.
As a candidate, I needed a way to stand out. A static portfolio was fine. A portfolio where you could actually ask me questions and get pointed to the right case study or resource felt different. I also needed to keep creating. Job searching can be draining. Building something real kept me in "do" mode.
I didn't know if a chat feature would be worth the cost or the time. I was literally about to be unemployed. Spending money on an API felt risky. I had to decide: ship something that might not "result in anything" or play it safe. I decided $5 was survivable. Anthropic's own limits let me cap spend at $1 so I could try it without fear.
The layoff made the website urgent. It also made the "should I spend money on this?" question real. Doing it anyway and capping spend was the compromise. And I wanted to prove to myself I could still ship: curiosity and being a do-er are what I pride myself on.
I didn't do formal research. I had a constraint (stand out, learn, stay sane) and an idea (chat that knows about me). I validated the cost in my head: $5 to start, $1 cap. I validated the learning: Cursor + Anthropic + streaming + Neon would teach me the full stack of a modern AI feature.
Use a single markdown context file for context.
I wanted to keep things simple. One file (about me, resume, tools, case study blurbs) is easy to update and enough for the model to stay on script and point people to the right case studies.
Cap Anthropic spend at $1 and add in-app rate limiting.
So I could ship without fear of a runaway bill. Rate limiting per IP keeps a single user from exhausting the cap.
Store submissions in Neon when DATABASE_URL is set.
I wanted to see if anyone used it. Optional Neon means the feature works with or without it; when configured, I get a simple log of questions (and optionally answers).
Design the bot to refer people to case studies and direct contact.
The goal was to get me a job. The bot's job is to surface the right work and suggest next steps (case studies, resume, contact), not to be a generic FAQ.
I went with a single context file: scope was just my own content, and I wanted to ship without over-engineering. I added summaries and pointers to case studies so the bot could recommend specific pieces.
A chat interface on my Nuxt site: hero CTA that goes to /chat with the question, a chat page with back link and message list, and a server route that streams from Claude using a markdown context file. Rate limiting and optional Neon logging round it out.
POST /api/chat streams Claude (Sonnet 4.5) with a system prompt and a single markdown context file.
I needed the model to know about me without building something complex I'd have to learn from scratch.
Server reads content/chat-context.md and injects it into the system prompt. Messages are converted and sent to the API. Response is streamed back via Vercel AI SDK so the UI updates in real time.
Per-IP rate limit (e.g. 20/min) and Anthropic spend cap so the feature can't blow the budget.
One user or bot could send a lot of requests and burn through the API budget.
In-memory rate limit before calling the LLM; 429 when over limit. Anthropic console spend cap set to $1.
When NUXT_DATABASE_URL is set, each user message (and optionally the response) is stored in Postgres.
I wanted to know if anyone used the chat and what they asked.
chat_submissions table: user_message, ip, created_at (and response if we store it). Insert runs when the request is handled; no auth, just a log.
To get started (Anthropic API key)
Spend cap so I could ship without worry
Time to ship
Retrospective: As I go through the motions of a layoff and think about who I am and what I want to do, the thing I pride myself on is curiosity and being able to figure it out. This project was a way to stand out, to learn, and to show that I'm someone with real curiosity who gets things done. It was $5, a few Cursor chats, and a decision to build instead of wait.