My Confession: How a 3-Year Developer Built an Enterprise AWS RAG Platform with AI, Curiosity, and Systems Thinking (Bonus Part 6)

Over the past five articles, we explored the intricacies of distributed systems, including SQS Claim-Checks, multi-format parsers, PostgreSQL pgvector HNSW co-location, Reciprocal Rank Fusion, SageMaker Cross-Encoder reranking, and Redis cache partitioning. Many readers may have assumed that the platform described in this series was built by a seasoned Cloud Architect with over a decade of enterprise experience.

However, I have an honest confession to make. I am only a 3-year veteran in the industry, currently pursuing my Master of Computer Applications degree. Moreover, I did not write every line of code in the repository. A substantial portion was generated, refined, and debugged through extensive pair-programming with AI tools, primarily ChatGPT, Gemini, and Claude (Antigravity).

Despite the common belief that AI can effortlessly build complex enterprise platforms, this is not the case. In reality, AI models lack innate judgment and rely on average practices found on the internet. These practices often lead to flawed RAG tutorials that are synchronous, single-tenant, insecure, and prone to hallucinations. While AI cannot design the entire platform, curiosity and systems thinking played a crucial role in guiding the architecture.

As a developer with limited experience, I navigated high-stakes architectural decisions by treating AI as a hyper-intelligent sounding board to challenge my technical hypotheses rather than an oracle. The core skill in modern software engineering has shifted from memorizing syntax and typing speed to systems thinking, problem framing, and architectural trade-off evaluation. When designing the ingestion pipeline, I asked questions about potential issues rather than simply requesting code for document upload.

Through this process, I learned the importance of understanding fundamental computer science principles, such as B-Trees, vector cosine geometry, thread pools, and database transactions. AI provided the syntax, but my domain framing directed the architecture. This approach paid off, as I encountered real-world issues during production deployment. For instance, PowerShell secret traps, SageMaker UTF-8 codec crashes, and BM25 zero-match blindspots required hands-on engineering debugging to resolve.