K,V caching mechanism on my whiteboard. The token under the dotted line represents the newest token If you’ve been following my previous blogs, you might have seen that I’ve been going on an LLM training journey. I started with building an LLM from scratch (training is still ongoing, by the way), and now I’m digging deeper into other topics such as K,V caching, Mixture of Experts, Reinforcement Learning, etc. ...
For the past few weeks I’ve been going through the motions of training an LLM. LLMs typically go through three stages: pre-training, fine tuning and alignment. I was having a bit of trouble fully understanding how Reinforcement Learning stage helps with alignment, and how it differs from the fine tuning stage, which is the reason for writing this blog. A few years ago, I was doing a contracting job as an ML engineer for an early stage startup. The founders had managed access to the brand new LLAMA 7B and 13B parameter models, and wanted me to fine tune it to better understand the capabilities of the model. ...
I load tested a Databricks SQL Warehouse as the backend for a high concurrency API, to find out whether you can skip the operational database. Here is what the numbers showed, and the more useful lesson underneath them.
This is the story of how I overengineered a software project because I made the wrong assumptions too early. This was originally going to be a success story of how I managed to rescue my friend’s website from his web developer who was keeping it hostage. I was going to write about how I used AI to rewrite a whole website in a day, despite me not being a web developer. In the end, I realised I could have done this in 20 minutes, rather than taking a full day. ...
Over the last 3 years I’ve spent a lot of time building highly scalable Data APIs, to enable teams to access data from different sources with minimal effort. There have been several mistakes made along the way, and many lessons learned. I believe these lessons will form strong foundations to build performant Data APIs from day one. Before diving into the lessons, I’ll define Data API. A Data API is an API service that serves some data to the user. The API backend connects to a database, such as Postgres, Redis or MongoDB, it runs a query based on the API call received, and it serves that data to the end user. ...
Over the past few months I’ve been using AI agents a lot to automate many engineering tasks that otherwise would take me much longer to do. I’m a firm believer that the era of writing code by hand has finished. With the right workflows these tools can do a lot of the work for us. AI is improving our productivity, and it will continue more and more as the tools mature. ...
Coding agents work best when they have a good feedback loop they can use to verify their changes. Without a good local environment and feedback loop, agentic development quickly becomes a waste of time. My first experience with software development was javascript, building websites using frameworks like React. I loved my development setup back then. I would always have a local server, automatically refreshing on every change. If anything broke, I would instantly find out and fix the issue right away. I would also have my unit tests running constantly catching bugs in near real time. ...
If you were to peek inside Postgres and see how the tables are stored, you’d be surprised they look nothing like the outuput you get when you do a SELECT query. In this post I look inside Postgres to understand how tables are stored in the filesystem. The goal of this post is nothing more than to understand a small part of the whole Postgres ecosystem to demystify things a bit. It is a step by step process to find out where tables actually live in the filesystem and their structure. ...
This has been one of the years where I’ve learned the most things, probably since I first started learning programming almost 10 years ago. I thought it would be a great exercise to put it into writing to reflect on it all. The Dvorak keyboard layout The first thing I set out to learn in 2025 was how to type using the Dvorak keyboard layout. The main reason was simply because it sounded like a fun thing to do. I still remember how much fun it was to learn QWERTY in school, going through all of those timed exercises. This time around it wasn’t as fun, and to be honest, it was very frustrating at times, especially when my brain completely forgot how to QWERTY and I was stuck typing Dvorak full time at very low speeds. ...
I have passed the AWS Architect Professional exam with a score of 791/1000! Preparing for the AWS exam I have recently completed the AWS Architect Associate Exam, and I thought the Professional would be a nice next step considering all the prep work I had done. I took about three months between passing the Associate and taking the Professional exam. For the Associate exam I spent a lot of hours each day going through the course content, writing notes and doing the practice exams multiple times, this time I didn’t have the luxury of time. I had to do all of the studying during work hours, where I could only spend about an hour a day (sometimes more, sometimes less), 4 days a week. ...
Today I wanted to write about something that happens to me at work all the time: the process of asking for technical help, helps me discover the answer to my question. For example, today I was stuck on some terraform config that wasn’t showing some of the changes I was expecting to see in the diff. I decided to start writing a message in a support channel. As I was writing the message I started thinking about whether I had covered all the bases of things that I could check on before bothering someone to help me on something I could have done myself. At that point, it occured to me that I hadn’t checked a specific file that might have some releveant information. So, I deleted my half typed message, checked the file, and voila! The reason my config wasn’t applying was because I forgot to explicitly add some variables. ...
Slowly changing dimensions, what a confusing name. This is the sort of term I’ve been hearing for a while now, but never really bothered to properly look up what it actually means. As it turns out, all it means is how you handle row updates/deletes in your data pipelines. Whether you decide to maintain historical data in your tables by appending new rows when a change happens, or you decide to update in place instead. ...
Interviewing for other jobs in tech can make you a more competent engineer. My premise for this argument is that when you prepare an interview for a new role you need to prepare to answer questions, technical and non-technical, in a domain that is different from your day to day work. This forces you to step out of your daily work and read and think about different domains and technologies. During the interview process you will get an idea of how well you performed, whether that is via self reflection, or feedback from the emloyer. This gives you a pretty good idea of where you are lacking knowledge and a good direction of where to focus next. ...
When working with distributed systems you can encounter terminology such as “replicas” or “standby” intstances. When I first started learning about distrubuted systems, I assumed “replicas” and “standby” were the same thing, but called differently when talking about scaling or high availability. While they can be the same thing, that is not always the case, so I wanted to clarify the concepts in this post. Replicas Replicas are one or more exact copies of the primary instance of your database. Their main purpose is to balance the load. ...
Practical strategies to protect your serverless applications from runaway AWS bills, including caching, rate limiting, WAF rules, and DynamoDB safeguards.
After ignoring the whole “agent” hype for a while, I started to get that inevitable feeling of being left behind. That changed when I came across this blog post explaining the basic flow of an agent. It made me want to try it for myself and see what it actually makes an agent. This post is my walkthrough of that process: step by step, showing how to turn an LLM into an agent. The code here is mainly illustrative, meant to highlight the concepts rather than provide a fully working solution. ...
I recently passed the AWS Solutions Architect Associate exam with a score of 848. This certification covers a wide range of AWS services which has taken a good amount of work and effort to prepare. In this blog post I wanted to share my experience preparing for the exam, sharing some info about the exam, and what has worked for me. Preparation I have about five years of on-and-off AWS experience, mostly with RDS, CloudFormation, Lambda, EC2 and a few other services. ...
Introduction This is a guide to myself if I need to deploy an Appsync App in the future. Appsync is a Serverless AWS offering for implementing GraphQL APIs. It integrates nicely with Aurora and DynamoDB. Other integrations are done via Lambdas. Architecture Overview Compontents AWS AppSync: The managed GraphQL service that processes API requests Aurora PostgreSQL: Serverless relational database for data storage AWS Secrets Manager: Securely stores database credentials AWS IAM: Manages permissions between services AWS CloudWatch: Handles logging and monitoring The flow of a request is: ...
I recently had to run load testing for an API that fetches data from Postgres. I was monitoring Postgres when running the first test, and I noticed that hundreds of connections were being opened. I was also seeing runtime errors in Postgres such as “no more connections allowed” or “out of shared memory”. Clearly the database was not prepared for the load. So I started to look for ways to improve connection management to the database, and I came across pgBouncer, a connection pooling service for Postgres. ...
In this post, guest writer Billy interviews Pablo on learning Dvorak. Pablo started to learn Dvorak in January 2025. This is an interview about his experience. First of all, can you explain what Dvorak is? Dvorak is a keyboard layout created in the 1930s, proposed as a more efficient alternative to Qwerty. Dvorak groups letters on the keyboard based on how often they are used (in the English language), so the most used ones are located in the “home row” and less used ones are located in locations that require further finger reach. ...