Portfolio - RAG Chatbot

The Challenge

Most PM portfolios are static pages listing past experience. I wanted to build something that demonstrates I can go beyond writing specs—actually designing and implementing an AI system end-to-end. The goal: a portfolio that itself is a product, showcasing RAG architecture, prompt engineering, and modern development workflows.

My Approach

Key Decisions

• →Chat-first homepage over traditional portfolio layout. The chat IS the product—if the AI experience is compelling, it proves the point better than any case study description.
• →pgvector over dedicated vector DB (Pinecone, Weaviate). For 25+ content chunks, PostgreSQL with vector extension provides the same capabilities without adding infrastructure complexity or cost.
• →Claude for generation, OpenAI for embeddings. Each model excels at its task—Claude's reasoning produces better conversational answers, while OpenAI's embedding model provides efficient vector representations.
• →Structured phases with autonomous execution. Breaking the build into 8 clear phases with quality gates enabled rapid parallel development without sacrificing code quality.
• →Evaluated retriever and generator separately rather than treating the RAG pipeline as a single unit. When answers are poor, the fix is different depending on whether the right content wasn't retrieved or whether the LLM misused good context—separating these surfaces the real problem.

Results

This project demonstrates end-to-end AI product execution: from architecture decisions and vector database design to a production-ready chat interface, all built in a structured, test-driven workflow.

Key Learnings

• ✓AI-assisted development changes the PM-engineer dynamic. Using Claude Code with specialized agents let me move from "writing specs for engineers" to "directing AI agents through implementation"—a workflow that will define the next generation of technical PMs.
• ✓RAG architecture is about trade-offs, not complexity. The core pattern (chunk → embed → retrieve → generate) is straightforward; the real skill is choosing the right chunk size, similarity threshold, and prompt structure to get useful answers.
• ✓Products differentiate through execution, not features. Every PM could list "built a chatbot" on their portfolio. Actually shipping one—with tests, security reviews, and production-quality code—demonstrates a fundamentally different level of capability.
• ✓AI products need evals the way traditional products need analytics. Off-the-shelf metrics like "hallucination score" disconnect from real user problems—starting with error analysis of actual interactions and categorizing failure modes produces evaluators that actually improve the product.

Tech Stack

Links