For me, 2025 was a great year for technical exploration, and iterating on a set of ideas making them more mature. Unlike 2024, I didn’t chase trends or get into shiny new tools. Instead, I focused on building systems that worked, at every iteration refining my approach to productivity, putting persistent effort and planning properly around the ideas. This retrospective is a catalogue of things that I have done over the year and learnt from them. I am writing down things and results more often as a regular exercise. It is also ok to have downtimes through out the year where I don’t need to produce anything.
Productivity improvements:#
- Switched to Linux as my primary workstation environment.Linux Migration . I also found functional replacements for all software I used on Windows and Mac.
- Switched to Hugo for the website and blogging more regularly. Hugo Migration
Personal Projects:#
- YouTube subtitle downloader & summarizer: The purpose of this tool is converting subtitles into readable markdown format that is catalogued for a Notetaking tool for easy reading on a phone or a tablet. So instead of watching a video, I read the content. I used non-LLM methods to restore punctuation and paragraph structure, avoiding the data and context loss inherent in LLMs. LLMs are stochastic in their outputs and often have guard rails that restrict certain content or words, leading to censorship or unintended summarization even when explicitly prompted otherwise. I did add LLM integration added to identify sponsored segments, self promotions, delete them from the text. I also added a score system for AI-generated scripts.
- Automated screenshot taking tool and monitoring website changes with vision LLMs.
- Energy price extractor and planner. I wrote a blog post on this Energy prices extractor . Remains AI free and with classic machine learning/rule based system.
- Dashboard for monitoring services on my home server.
- Street photography tagger with a vision LLM running locally. I wrote a blog post on this Street Photography Captions
- Posture detector and corrector using three web cameras from three positions.
- I wanted to write a Music player to replicate Cowon/JetAudio music player feel and functionality. I wrote a few iterations in python as a learning exercise into audio processing. These interations are not meant for performance though it is an interesting exercise to see a music player run 100% on a modern AMD 16 core CPU with all the bells and whistles. Rewrote it in Rust for performance but it does not yet have the same amount of features as in python. The Music player will remain AI free.
- Real estate scrapper/tracker/alert system for Den Haag and nearby Gemeentes.
Incomplete and to be completed in 2026:#
- A bird swarm detector with webcam. Target birds are the Hague crows and if I am lucky maybe parquets (they are usually too fast for the webcam that I am using).
- Interactive decision making tool when having multiple paths for a given task by weighing in all the factors that go into the decision.
- Breathing and mindfulness detector with a polar H10 band integration for adding breath and heart beat information. I want to integrate this into my music player from above.
- Downloading and making time lapses of satellite weather pictures. It is currently as a data collection exercise. As a final product, I want to study a whole year for one country (India) and build a classic vision model to categorize different weather events on a long time sequence.
- Index fund tracker and index fund picker.
- Learning and programming on a NPU with OpenVINO.
Explorations:#
- Running an on-demand small LLM on a sub 75w GPU 24/7 for a month scheduling and handling queries according to energy prices.
- Going through a month without using LLMs to reset usage habits. LLMs are good as occasional tools but not for everything that we do as knowledge work.
- Group of LLM experts discussion. I wrote couple of blog posts on this Building a digital twin of a Virtual Research Lab Building a digital twin of an Investment Committee
- A LLM summarizer of each chapter within a book and then comparing this with the actual book. It showed me how terrible in general the usage of AI summarizer is in SciFi and travelogues genres because almost all the nuance is lost. Built a MCP server around it. It is no longer in operation for the above reasons.
- Studying heat absorption into oceans and their spreads over the last 3 decades.
- Studying heat transmission from solar cycles and their effect on earth weather. Solar Irradiance
Homelab:#
- Running a Homelab 24/7 (99.9% uptime, 4 months) with minimal energy usage. Homelab 24/7
- Pruned and dropped services as part of an exercise in minimalism.
- Not using any vibe coded self hosted services in favor of mature alternatives.
- Built documentation and setup scripts to enable quick recovery if anything failed.
AI:#
- I stopped trying out most fancy new AI tool that wont survive another year instead focused on working with mature tools and agents that I had more control and experience over.
- I think Deepseek OCR is the best AI tool from 2025 that I use on a regular basis Deepseek OCR . It is a fundamental shift like [markitdown] (https://github.com/microsoft/markitdown ) when it got released in Dec 2024. -llama.cpp is much faster and reliable than Ollama.
- Qwen 3 are a great family of models for thinking.
- I still prefer older llama models for a lot of JSON processing because of their reliability.
- GPT OSS is also a decent model but Qwen 3 is a bit better.
- I completely migrated away from buggy Gemma text models.
- I also decided not to use any of the AI browsers due to significant security risks and lack of data protections.
Github commit graph for 2025#
