Jinwoo Je | Portfolio

The Architecture of Understanding

March crystallized a fundamental shift in how we build and learn. The old paradigm of mastering low-level details before tackling higher-order problems is being inverted. Now, understanding the architecture – how pieces connect – matters more than understanding each component in isolation.

Cloud Platforms: The Abstraction Revolution
Learned the profound difference between infrastructure-focused platforms like AWS and developer-friendly platforms like Vercel and Supabase. AWS requires deep infrastructure knowledge – it's built for enterprise teams with specialized roles. Vercel and Supabase abstract away complexity, enabling individual developers to move at startup speed. Different pricing models reflect different value propositions: AWS charges for resources consumed, developer platforms charge for problems solved.

Systems Thinking Over Detail Mastery
The bird's-eye view beats deep dives when you're building full-stack applications. I've been narrowly focused on Python for time series work – 90% of my coding time. Building ZipScape forced me into JavaScript and Next.js. The lesson: learning the right tools for specific jobs accelerates progress more than deepening expertise in familiar tools. Breadth creates more leverage than depth when you're working alone.

Gemini 2.5 Pro: The New Standard
Gemini 2.5 Pro has become my default model for most tasks. Blazing fast, unlimited context, shares thinking tokens – the complete package. DeepResearch helped me complete a project I'd been thinking about for years. The image generation capabilities are remarkably coherent with strong object consistency. Second month on Gemini Pro subscription, and it's proving its value consistently.

AI Integration Across All Fronts
Deliberately used LLMs across every available interface: chat, integrated IDEs, thinking models, deep research, MCPs, CLI agents. First time maxing out my Cursor subscription (500 fast requests) and Claude Code credits ($30). This wasn't experimentation – it was systematic adoption. When tools work, use them fully rather than cautiously.

Beyond One-Shot Prompting
One-shot, ambiguous prompting fails beyond toy use cases. The magic happens in agentic workflows with proper planning, design, and systematic prompt management. Templates and specifications matter. Research becomes essential – you're acting as PMO for a team of AI experts. You must understand the problem space, distill required components, and build effective prompts. It's project management for artificial intelligence.

The Three Pillars: Model, Context, Prompt
Strong model + appropriate context + clear prompts = reliable results. This isn't just best practice – it's the only practice that works consistently. Each pillar supports the others; weakness in any area undermines the entire effort.

MCP Clients: The Underexplored Opportunity
While MCP servers get attention, MCP clients represent the bigger opportunity. Claude's desktop app is solid but limited to Claude models. Cursor works well but is coding-specific. ChatGPT announced MCP adoption but hasn't delivered the desktop client yet. Really hoping Gemini adopts MCP and creates a desktop client – that would be transformational.

Step-by-Step Prompting Strategy
Prompting works best when broken into steps. Avoid one-shot prompts like "build me an app that does..." Document prompts and iterations during development. Still developing better prompt management systems – currently keeping notes, versioning prompts, tracking iterations. This meta-process matters as much as the technical work.

ZipScape Lessons: The Cost of Moving Fast
Building ZipScape taught expensive lessons about Git discipline. Lost a weekend's work due to a Git mistake where .gitignore files were deleted and couldn't be recovered. Learned to be cautious in agentic mode – understand what the AI is doing and course-correct when necessary. Deployment complexity is fascinating but unforgiving: configuration, environment variables, build processes, package installation all must work together seamlessly.

AI Project Initiation Anti-Pattern
LLMs aren't good at initiating code projects from zero. Much more success setting up the environment and basic folder structure first: uv init project, pnpm create-next-app, then handing off to AI. The AI excels at building on foundations, not creating them from nothing.

Test After Every Step
Manual testing after each code change is time-consuming but essential. Currently running servers and testing personally after changes. Need to learn programmatic testing – creating tests for each prompt iteration. This is the next frontier for AI-assisted development.

Clarity as Competitive Advantage
Knowing what to do and what not to do becomes crucial when working with AI. Clear constraints and explicit requirements prevent AI from optimizing for the wrong objectives. Precision in communication translates directly to precision in output.

Big Data Reality Check
First encounter with a dataset too large for conventional loading methods. Three days attempting to load parquet files through vertical scaling failed completely. Required distributed processing approach, but lacked IAM permissions. Created workaround through filtering and direct ETL processing. Planning to learn partitioning, EMR clusters, Hadoop, Spark by end of Q2. Sometimes you hit the limits of your current approach and must level up.

Business Skills Matter More
Business acumen, product sense, and critical thinking increasingly important compared to pure coding ability. LLMs abstract away technical implementation details, similar to how high-level languages replaced machine code. The future belongs to those who understand what to build, not just how to build it.

Naval's Wisdom Applied
Key lessons from Naval Ravikant: True intelligence means achieving desired life outcomes by knowing how to get what you want and wanting the right things. Avoid autopilot living through deliberate decisions. Spend significant time on major life choices. Balance optimism with quickly cutting losses. Use objective reasoning rather than self-limiting labels.

Travel Bonus
Alaska Airlines launching direct Seattle-Seoul flights in September. Personal win – they're my favorite airline, and this route eliminates connection complexity for Asian travel.

March's overarching theme: architecture beats implementation, systems thinking beats component mastery, and business judgment beats technical skill. We're witnessing the emergence of AI-native workflows that require fundamentally different approaches to learning, building, and problem-solving. The future belongs to orchestrators, not operators.