Pre-Core 119 Updates Before Migration
Migrating previous blog posts here starting from 2026-06-16 up to 2026-07-20:
2026-06-16
Overloading My Plate
Been focused on reviewing all aspects of the general field of software engineering and specifics of the Java programming language and related technologies on the backend using Spring Boot for REST API and with Spring Data JPA, Identity Platform, Security, Observability, CI/CD, Cloud, Enterprise Integration, Middleware, Design Patterns, and others.
I got rusty on the algorithm side so I started just recently on the basics of algorithm and its code to awaken my neurons. I started doing this using the Java programming language but I realized it is better to do it recently with the C programming language since both are the same in the code syntax and almost the same in the code semantics except for things like array parameters where in the latter it becomes a pointer.
Recalled that I installed C toolchains both my smartphone and on my desktop but I forgot which tools exactly so it took me some time to get the information for it. My smartphone has Termux for a Linux emulator and installed only the Clang compiler there and was done so I can make my assignments via my phone back from last year while was under the BSEE program. In my desktop I installed Mingw-w64 UCRT64 with GCC for Windows but yesterday I added to install Make and it was the first time I was able to run Make as part of the C toolchain specially when I started splitting my algorithm C programs into different areas with different header files and source files. Make simplifies the complex compile, link, and build process into fewer command line commands such as ‘make clean’ and ‘make’.
At the same time, I just paid for a Udemy training on AI Engineering so I can speed up my transitioning to AI Engineering as all my existing Software Engineering skills are already part of AI Engineer and what is missing for me is Python, ML fundamentals, LLMs and Generative AI, Vector Databases, AI Frameworks like LangChain, LlamaIndex, Haystack, Model Serving and Inference, and then RAG Architecture.
Also miss updating my blog to keep make memories and keep track of ton of things in my life.
2026-06-16
Historic Moment for Alfred
Today is a historic moment for Alfred, my workstation PC because for the first time a local instance of an AI Model is being downloaded and being prepared to run on Alfred.
Just a small AI Model for learning AI Engineering, llama3.2:1b. I can install llama3.2 which is a bigger than the former but I do a lot of other things in my PC hence I still need room for computing resources hence the smaller llama3.2:1b is suffice.
2026-06-17
Git Clone, Microsoft C++ Build Tools, Cursor, and Back to Algorithms
Cloned the git project that I will use for the AI Engineering course from Udemy and had to install the MS C++ Build Tools and then enable the long filenames in Windows 10 Pro. Finally installed Cursor and I didn’t know it is based on VS Code.
Right now I am back on the search and sorting algorithms sharpening my coding skills for the different algorithms.
2026-06-17
Studying LCS Algorithm
Why I have not seen this Longest Common Subsequence algorithm back in BSCS and in MSCS? This is related to Palindrome, a string operations algorithm.
2026-06-18
Python, Ollama, Long Common Subsequence Algorithm, and Industrial C
Last night I was on the part where a Python lab needs to connect to an OpenAI API Key so I went the alternative for no spending hence I need to work on the labs so that the Python lab runs on Ollama instead without a need to pay for an API Key.
At the same time I was studying the implementation of LCS algorithm both in Java and in C. The Long Common Subsequence Algorithm is related to the string operation algorithm Palindrome where the string is reversed. The Java implementation is a bit cleaner than the C one.
I’ve started seeing how an Industrial C program is written. Not much different from Java as there are guard checks for say nullity of the function parameter or checking for empty string. Another is the use of libraries where functions such as strlen is taken from and the size_t type which has advantages unlike the typical int type.
Peculiar is the absence of palindrome operation functions such as reversing a string but it turns out that the culture in C is to no bother making a library for trivial coding unlike in Java where string reversal actually exist in the class StringBuilder.
2026-06-25
Weather Service Exercise
Did a weather service exercise and stepped it up to production level based on the requirements although I was tempted to add observability but I submitted the codes without it and planned to add observability dependencies in Spring Boot and tag it as another version.
A lot of the traditional best practices which were manually done before, I have included as dependencies and forces the developer to stick with it because failing to follow the best practices with these dependencies breaks the Maven build so there is no leeway to not follow the best practices.
One of the headaches is adding the security in the quality gateway because NIST’s update service for the vulnerability database for my local keeps failing. I did increase timeouts and it pushed the updates a bit further but it still fails at times because the NVD API update services keeps failing at times and so say not even reaching 25% of the total updates downloaded and then the services fails.
I will be back later for some technical reviews again and so AI Engineering course get’s sidelined for a bit while MIT course 6-5, and 8 is still out for long.
2026-06-30
WeatherService Release
Released the WeatherService today.
Had to replace the OWASP Dependency Check with Sonatype because it took me many days with updates from NIST to my local NVD database taking at least 8 hours before it hits some kind of timeouts or error.
Did played with timeouts and retries for OWASP Dependency Check until it seems to be working fine without timeout or retry issues but then when it approach 99% suddenly the exceeding varchar (1000) constraint was thrown and basically ended the update.
Sonatype didn’t have those problems it was fast to scan my project for CVE vulnerabilities.
2026-06-30
AI Engineering and The qwen2.5.-coder:14b Open Source Model
Juggling the AI Engineering course with other things and though I am not fulltime on the course, I been learning Python very quick including interfacing with AI Models using the OpenAI interface including directly interfacing with a REST API endpoint.
Also realized that there are two kinds of prompts with different use, system prompt for setting the context of the interaction and the user prompt where an actual problem, a topic, or task is discussed.
A good understanding between the difference of the frontier models and the open source / open weight models was imparted from the course and also got familiarized with the different systems in LLM starting from a base model, to a chat structure, to reasoning system, and to a hybrid of chat and reasoning system.
Subscribed already with OpenAI’s Codex as I am very much satisfied with the smartness of this AI coding assistant which is a combo of an AI Model and an AI Agent tied up in IntelliJ.
Got the following AI models under my local instance of Ollama:
PS ollama listrojects\llm_engineering> NAME ID SIZE MODIFIED PS C:\opt\ai-projects\llm_engineering> ollama list NAME ID SIZE MODIFIED qwen2.5-coder:14b 9ec8897f747e 9.0 GB 41 hours ago qwen2.5-coder:7b dae161e27b0e 4.7 GB 2 days ago llama3.2:latest a80c4f17acd5 2.0 GB 12 days ago llama3.2:1b baf6a787fdff 1.3 GB 13 days ago
and tried setting up VS Code with llama3.2:latest, qwen2.5-coder:7b, qwen2.5-coder:14b but I am not satisfied. The only consolation is that I have proven that my machine can run a 14 billion parameter AI model but slow. In terms of coding, even with Continue and Cline extensions in VS Code together with the top of the line local AI Model that I have which is qwen2.5-coder:14b I still didn’t like the performance. I can tolerate the slow response but the smartness is nowhere near a sober ChatGPT 4.x or ChatGPT 5.x or even Codex. Seems like the performance is the same as a hallucinating ChatGPT 4.x and Github Copilot from last year is even better. Only good for experimentation.
2026-07-05
Another REST API, VS Code Insider, and Working on an new AI Model
I did another REST API and submitted it last week and like the WeatherService REST API, it has the same quality gates setup. The only difference is that while the WeatherService is a demonstration of calling a downstream API with caching and circuit breaker, the new service is synchronizing some data from different sources where one downstream is served via WireMock, another is a drop file. Again, it also has circuit breaker.
On another note, I been busy with my AI Engineering course and setting up an IDE with a local instance of AI models for agentic tasks and nothing made it acceptable. VS Code with Continue and Cline extensions and connecting to my local qwen2.5-coder:7b and 14b impractical to use but when I gave qwen2.5-coder:7b and 14b a task as a standalone LLM via Ollama, the response time was acceptable and the coding problems were solved with ease.
Eventually I settled with VS Code Insider and it comes with a GPT 5 mini AI agent in the form of raptor.mini. The best part there although it is free but with a cap, the credit units are a fraction which is lower than the ones offered from frontier models.
Was also looking for ways to run AI bigger AI Models on consumer devices such as laptop and a smartphone and initially I was thinking of llamaC++ and AirLLM but I deferred the plan for it indefinitely and decided to look into the problem of transformer architecture which led me to implement an experimental AI Model. Still a playground and at the early stages for now.
2026-07-20
Week 2 & 3 in AI Engineering, Mastodon Instances Anomaly, Optimized Local AI Model Setup, and Local AI Model Internet Access
The final laboratory exercise for week 2 under the AI Engineering course I made three-way chat with local LLMs running via Ollama. Initially I tried a four-way chat which includes me, GPT-OSS:20B, Qwen2.5-Coder:14b, and llama3.2 which is 3b and it started in heavy disk pagination while also taking lots of time before the 3rd model responds in the chat. This is aggravated by the caching tool that I used (LiteLLM) throwing a timeout exception and since the situation is not pragmatic anymore, I mean who wants to chat with 3 AI models running locally if it’s slowing things down? So I decided to remove llama3.2 and the chat latency becomes tolerable. Had to enable latex rendering feature in Gradio so the chat front-end I am using to display math answers in a more human readable form.
Proceeded with week 3 and already noticed how it is very different from week 1 and week 2 in which OpenAI API is used to connected to a locally hosted AI Model via Ollama running on my AMD Ryzen 7 5700G with an integrated Vega 8 Radeon GPU. In week 3 the laboratory is not a local machine running Python sources using Jupyter lab notebooks but a cloud hosted laboratory running a Tesla T4 GPU which is awesome! The approach also is not via OpenAI API but via HuggingFace Transformer library with HuggingFace Pipeline using Pytorch.
Codes there were run in the Tesla T4 GPUs are using Diffusers to generate images via text-to-image AI models in which the number of steps was set from 4 to 40 before the image can be generated. Another is using text-to-speech.
The insights from week 1 to 2 and in week 3 although not yet finished is that high level access to AI models at least on text only prompting can be achieved with higher abstraction using OpenAI API. Fine grained control on AI models that are open weights / open source can be done using HuggingFace Pipelines. So while the former connects to an LLM that is separately hosted, the latter embeds the AI model in the python source with fine grained control.
On the other hand, my Mastodon account in the Qoto instance went down with Qoto. I noticed that one of the founders and admin of the instance has last toots since last quarter of 2025 and never heard from since then. Coupled with that observation is the 0 trending topics in the Qoto instance which made me suspect something is wrong with the instance. Regular http errors during posting on the first try or sometimes even on the second and third attempts before finally getting my toot posted. Lately Qoto went offline and never came back and so is my account. Had to move to another account on universeodon but then just today it went down also. Coincidence that two instances in Mastodon where I have accounts went down recently. I don’t believe in coincidences, it’s mostly an anomaly.
Since my work on the laboratory exercise for week 2 where the four-way chat was aggressively making my disk do pagination, I noticed that while the loaded AI models are active, meaning they are working on some problem I gave them, my CPU usage tops at around 100% yet zero activities on Compute 0 and Compute 1 of my Vega 8 Radeon GPU. So while the CPU is topping, my GPU is just hitting 2% max at some point. The last straw was when I tried Microsoft’s Phi4-reasoning:plus, it took 30 mins and yet it was still working on a calculus integral problem.
Decided to use alternatives to Ollama and some quick research and with the help of AI, I decided to use LM Studio. Turns out LM studio with Vulkan in llama.cpp is a match for my Vega 8 Radeon GPU hence, local LLMs hosted in LM Studio in my local is able to offload to GPU while back at Ollama it only supported Cuda which means it is for NVIDIA GPUs only. Now, local AI models when actively working on some problem is only topping my CPU at around 55% while it hits max at 32% on GPU at times and then consist readout on Compute 0 with high graph trends. The AI models are also quick to solve problems or respond to my prompts.
Finally, I managed to setup LM Studio with web access using duckduckgo so that my local AI models in LM Studio can search the web whenever they need.
2026-07-20
Research Work on AI Models and Earthquake Forecasting
Recently I did uploaded to my Github an experimental C program for a non-transformer architecture based LLM but it’s on a very early stage. Then it was followed by Elon Musk claiming that AI models can actually create the executable binaries without a programming language source.
I got curious and did some research. Indeed, AI models can actually create executable binaries directly without a human readable programming language but the problem now arises about the verifiability of executable binaries in terms of functionality, non-functional aspects, quality, and security.
Furthermore, an AI model will have to generate an intermediary stream of binary 0 and 1 before a final executable binary is created and that stream of binary 0 and 1 is so much larger than a counterpart programming language source.
An idea came up and a paper was produced which was later submitted as preprint open-access paper to Zenodo. Next steps is to submit it to peer-reviewed journals.
On a separate endeavor, I been working on a novel aggregate of mathematics for an earthquake forecasting framework. Although the mathematics are aggregated from several sources, it is combined in novel ways in order to forecast earthquakes much effective than what we currently have.
The work is currently pending USGS catalog shadow forecasting after a number of parameter calibration cycles where done.
Best case would be to publish the foundational mathematics including the parameters but that might be harder to do so at the least, the foundational mathematics should be published instead and have other scientist do the appropriate work on parameter calibration instead.
This site will document work in engineering, computing, artificial intelligence, physics, Linux, and experimental technology.