If 2023 was about wow, and 2024 was about adoption, then 2025 has officially been the year of building.
Sitting here on the last day of the year, I realized my entire relationship with AI has shifted. Last year, I was happy just typing a prompt and getting an answer. This year? That wasn't enough. I needed to know why it worked. I needed to crack open the black box.
This is the story of my 2025—the year I stopped just "using" AI and started actually engineering it.
The Shift: Knowing "What Happens Behind the API"
The biggest unlock for me this year wasn't a specific model; it was understanding the architecture. I moved away from simple "input-output" thinking to understanding Agentic Workflows.
It’s one thing to ask an LLM to write code; it’s another to understand the orchestration behind it—how the model plans, executes a tool call, observes the output, and iterates.
Two tools that defined my 2025:
LangGraph: This was a game-changer. I finally moved past simple chains to building stateful, multi-actor agents. Learning how to manage "state" and cyclical graphs allowed me to build bots that don't just hallucinate an answer but actually reason through a problem, check their own work, and retry if they fail.
Windsurf / Cursor (The Agents): These aren't just autocomplete anymore. In 2025, tools like Windsurf became my pair programmer that could see my entire codebase. But the real learning happened when I dug into how they index the code using RAG (Retrieval-Augmented Generation) and embeddings. Understanding vector databases wasn't optional anymore—it was essential.
The Hackathon: From Idea to Top-10
The highlight of the year was undoubtedly the national hackathon. It was the first time I built an end-to-end AI product—not just a wrapper, but a full-stack solution with its own reasoning engine.
We were exhausted, fueled by caffeine and debugging errors at 3 AM, but seeing our agent successfully navigate a complex workflow without human intervention was pure magic. That project didn't just run; it soared, landing us in the Top 10 nationally.
It validated a massive lesson: You don't truly learn AI until you ship it.
The "One Tool a Month" Habit
To keep up with the insane pace of AI in 2025, I made a promise to myself in January: Learn one new tool or flow every single month.
It sounded daunting, but it became my anchor. One month it was multimodal processing, the next it was fine-tuning small language models (SLMs) on edge devices. This constant drip-feed of new knowledge prevented me from getting overwhelmed. It turned the "fear of missing out" into the "joy of figuring it out."
Don't Forget the Basics (Revisiting Old Notes)
Go back to your old notes.
Somewhere around October, I got stuck on a complex problem. I was trying to use the latest, fanciest framework, but nothing worked. I opened my notes from 2024, went back to the fundamentals of prompt engineering and data structuring, and found the solution in minutes.
Sometimes, in our rush to learn the "new shiny thing," we forget the foundational principles that actually make these things work. Your old learnings are the bedrock that supports the new skyscrapers you're building.
Onward to 2026
If 2025 taught me anything, it's that the technology will change, but the hunger to learn is the only constant. Whether it's dissecting a new agentic framework or debugging a messy graph, the goal remains the same.
Top comments (0)