I had focused on tech on this reviews. I would keep it same this year, I might write a separate non-tech, a human yearly review on my separate blog.

Gist

I moved from a software developer to a reliable and responsible developer. I took ownership and sat on problems more than the level of my comfort zone. It made me a better developer, it helped me sympathize the people around me, and connect more. These moments might create life long relations, but from the macro perspective they looked frustrating and annoying at times. Here I am still finding peace and standing and walking through what has not been a smooth journey both at work and life. I was thinking work and life could mean same thing once I got out of college, but those kept diverging and I had to keep both my feet at par with each one's pace.

I hit you with quotes and I will again

"Its easy to judge anyone, its harder to be someone"

I have read a lot of books, and one quote just hits me is this, I will explain this in my non-tech review post. Its quite easy to point out flaws and judge the other person from outside, but the same person if you were in his or her shoes, its quite hard to the otherwise.

It's easy to judge a developer, but people slowly realized vibe coding hits a halt at some point, you are not replacing a developer, you are trying to replace someone who cares deeply about the problem and not someone who just can write code.

Vibe coding rose and fall with the same speed.

Key Achievements

  1. Got promoted as a Software Developer from Junior (completed 2 years at Docsumo 8 month internship + 16 month full time)
  2. Wrote 52 newsletter editions consistently reflecting on weekly learning

I didn't finish any goals I wrote in 2024 review,

  • Write blogs for FreeCodeCamp (at least 3) (none written)
  • Learn Kubernetes in detail (not yet)
  • Write 25 articles on 100-days-of-Golang (wrote 3)
  • Learn Rust or Zig (Nope)
  • Get 2K subscribers on YouTube (Nope, not even 100)

I want to straight honest here, things change rapidly in 2025, I barely had things in control. I was consistent, I had time, yes! But I had to pivot to other goals to keep my self sane.

Not clearing 5 goals doesn't mean I cleared none

I did

  • Understand AI Agents before the hype (early January)
  • Learn in depth about SQL and SQLite with a learning log
  • Complete Advent of SQL (15 blog posts)

Also some broader or abstract goals like

  • Adapt LLMs in the development process
  • Reflecting more than consuming
  • Understanding one technology in depth (sqlite)
  • Complete things rather than shove it off after a while (project graveyard, or irresponsible employee)

Let's break it down to 3 things

  1. Work
  2. Side hustles
  3. Learning

Work

I work at Docsumo, so I work with documents. The field of intelligent document processing is taken by the storm by LLMs. Prior to LLMs, there were ML and rule based systems to extract data. Now we have LLMs, we not only can extract structured data, but can have a conversation with the document.

It has changed the way document are treated. We decided to go full on LLMs and use the latest model for extracting data. It was tough, there were issues, wired behaviors and of course hallucinations. Most of them were related to prompt but also more on how we call the LLM, through plain api, with structured output, or as an agent.

We started with the first simple api call, sat on it for quote some months, but external factors forced me to use other tools to find the extraction more accurate. I started experimenting with LLMs more than ever from the 2nd quarter of 2025. LLMs were released every week with minor improvements or major shift in tool calling. Everything had to go through the LLM, write and review code with Coding agents, use LLM to extract data, debug with LLMs.

There was a tremendous shift in how we develop software and I was blessed and lucky to be in a company that helps developers stay up to date with the cutting edge tools like Cursor, Windsurf, Code Rabbit, Gemini, OpenAI playground and what not. I was given these tools, as a new toolkit. No one actually knew how to use them, they knew the general sense of intuition, but it was like a alienated tool. Everyone gave their own shot at using it, and most them hated it at first, but then they understood the pattern, changed the way I and they used it and almost in no time it became the default editor and tool that we use now.

I never new change in software was that rapid! Maybe this is a special time, but every time we say that.

I honestly was reluctant to use AI tools as daily driver at first, but saw the speed of others and it bogged me down that I was missing out on something. And that was code-review and hands off understanding of the code.

It was very hard for me to understand and debug the code that I haven't written, I just accepted it without understanding, because it feels daunting to review code. Especially when you haven't written and reviewed a lot of code, human code. Its almost like reviewing the classical books (from Shakespere and Tolstoy) after reading Harry Potter. You don't have the vocabulary to understand those books, you can't review things that you don't know the patterns and the nuts and bolts of. It is possible but its quite tedious and overwhelming.

"Overwhelming" this is a underrated term for what I want to describe how I felt working throughout the year. Coding was supposed to be hands on, but when exposed with a magical bot that can write and churn hundreds and thousands of changes in few minutes is just not processable from my tiny brain.

I failed badly at using these tools, with skeptism. But then slowly over couple of months, with the right task coming at hand, and trying to throw anything at it, I understood. What can work and what can't work given a task to an LLM. I realized the hard way, that LLMs are good for throw-away code. The code that you can just use it for producing something that you don't care about how it did it. Like log analysis, scripts for downloading data, filtering files based on some names, analysing and combining bunch of CSVs, I don't care if you use proper variables and correct abstractions to do those things. These are the things where I never second guessed myself to throw at LLM and take whatever it spat out with a grain of salt and modified it till it made sense.

I am still finding it hard to write everything of that sort to LLM, as nothing is left to do! I feel bored! Coding was supposed to be fun, a bit time consuming but a good exercise to the brain. But people have started preferring velocity over craft. I am not against it, but sometimes we get too carried away with human ambition.

I hope we slow down in 2026, though I am skeptical about it, it might just get worse.

We need time to understand and explore what can be do with the current set of tools and technologies that these AI labs and all of us have created and release just in the last year. There is absurd amount of things that we need to build with them, in fact I think we can stay as is without new models or breakthrough for decades with just these tools. But Damn! We are not stopping here, are we?

I think in 2026, I need to build more and get a taste and expertise in systems. I am an engineer and writing code is already obsolete. The only thing that can stand out is understanding how things can fit and change in existing or new system software.

Just technical metrics, I don't care but gives me a sense of wonder of wow! that was quite a lot!

  • 336 commits on core services that I maintain
  • 101 PRs closed
  • Favorite model? GPT 5 (I am surprised cursor)
  • 708 M tokens burnt (geese)

Side Hustle

Side projects of course!

I had 100+ ideas but never picked up anyone of them seriously. What is wrong with me?

Is it due to AI, no. It was same as before AI too. AI just made it easier to reach that point.

With LLMs and these CLI coding agents, I can scaffold projects in seconds than in minutes. I initially thought it was cool, so that I can actually build the core thing, once the boilerplate code is created by LLM. But boy! I was wrong. It creates a layer of understanding that I then have to dread through, then with the half mine and half LLM baked system, I have to steer myself or the LLM in completing the project. And that's the bottleneck I am not able to solve it yet. I am writing this and realizing the problem now. Reflection is key in understanding problems. Writing is the easiest way to it.

I don't have much to say, I tried building CLIs, TUIs, Web apps, all failed due to my non-thinking of the parts, and was too lazy to write myself or LLM was not able to understand me, my prompting skills have gotten better over time, but I am bad at expressing things in plain english quickly, I need time, but the excitement of seeing it dampens it.

Side project graveyard is ever expanding, and we need something to cure it! LLMs have made it easier to expand, but also they have to power to complete them, the skill is the gap that is preventing me from achieving just that.

2026, Fingers crossed. I can solve at least one problem that people can relate to and love.

Learning

This I think is the least interesting thing for people right now, LLMs have made it look obsolete to learn what's new in SQLite 3.51, or how to use f and t strings in python 3.14. But I think it had given us ample reasons to learn anything. Why? Developer are paid to fix problems, not write code. If you know certain things, you always and always will have advantage over LLMs. You can be smart by leveraging LLMs to learn, in the thinking mode, it can spit out legit facts, but do the opposite thing of what it intended to do, you as a developer can see and do things better than LLMs. This is still a problem in most of the overthinking LLMs. They are getting better, but a human thinking will always out-power an LLM with fake thinking.

I decided to learn SQLite in 2025, and will continue to learn in 2026, its something that cannot be completed. I might also start learning Kubernetes the same way, though I wanted to learn it since 2 years but never started.

People will soon realize that learning is something that they can't outsource to an LLM, nor the thinking. Because humans at core are curious creatures, if you think you can live without learning anything, than you are not a software developer. Simple. Learning is an lifelong process, nothing can stop it, it could be never completed. There will always be something to learn. For instance, have you ever wondered how LLMs can read files? Like PDFs and words or images, since the data is not a clean text, how does it parse it? Well that is a rabbit hole for learning which could benefit you over just using LLM to churn and do a trial and error for correct result. If you have a tool, you need to understand it to the finest to hone and master it, for no such tool can be mastered.

Learning to learn is going to be more important than building in the future

That's just my opinion! Who knows, human just take AI slop and feed it to the next AI until they have something they like and never learn! That would be the worst of time for humans like me to live.

  • 52 Newsletters published at every Saturday midnight IST
  • 53 SQLogs written
  • 3 in-depth posts about Golang
  • 458 commits
  • 209 day reading streak on daily.dev

Lessons Learnt

  1. Never outsource thinking
  2. Be with people and speak out
  3. Habits are actually easier to build, the temptation to skip it for today is what gets better of me
  4. Developers were never for writing code, they were for caring about the problem the deepest way
  5. I can learn anything, the only thing in my way is my mindset and distractions

Keeping mind peaceful from outside though a lot is burning in the house, is something I juggled with.

Off to 2026

There are surley certain things I want to do explicitly in 2026 like

  1. Understanding databases on a deeper level
  2. Understanding LLMs on a deeper level
  3. Building tools with LLMs for fun
  4. Writing on my blog with less friction to publish thoughts, journals and devlogs or TILs
  5. Building educative material related to databases and SQL

Yes, these are very specific, but I don't like lame and broader goals like "Be a better developer", "Understand systems", and what not. I want to dive deep, I want to replace scrolling with action.

2025 was wild, it was slow yet models kept falling like dog water!

I don't expect for LLMs to go in 2026, that is the last thing that could happen, but I can see the thing starting to build around them, tooling for understanding them, helping with the smallest of things human care for, its already in all the things we use, yet people don't use it as reliable tools. The ones they do, they might soon realize what kind of fire they are playing with.

Anyways! I hope everyone a Happy New year and may 2026 be kinder, more fulfilling to you.

Keep Thinking!