A visual, ground-up walkthrough of RL for LLMs — how GRPO turns a group of rollouts into a gradient, why the absolute reward never matters, what reward hacking looks like up close, and how to train a reward model from nothing but human rankings.
How supervised fine-tuning actually worksA visual, ground-up walkthrough of supervised fine-tuning for LLMs — what the loss actually measures, why the off-by-one trips everyone up, what masking really does, and how much of this you can do on one GPU.
Building a watering system for my balconyA weekend electronics project to keep my Berlin balcony's plants alive through the summer heat — an ESP32-controlled drip system, the components, the wiring, and what I'd do differently next time.
Reflections from my time at HubSpotAfter 2.5 years leading engineering in HubSpot's Flywheel Product Line, I reflect on turning around an org of 20+ teams and the lessons I learned building internal AI products for GTM.
Thew new talent bar for software engineersWeb software engineering is changing rapidly. While this shift is not equally distributed and is often easy to overlook in the day-to-day, it is coming whether you are ready or not. The reality is that writing software is becoming cheaper...
Lessons from Intercom on being a product engineering leaderI am finishing my time at Intercom and this post is an attempt of reflection and summary of lessons from my time there. I had a privilege to work there for 6 years, across 3 roles, in 3 product groups,...
Turning around a group of teamsWhen you realise that the org you are leading became dysfunctional, it can feel hopeless. The role of engineering leader is to identify, diagnose and improve. Here is a framework that I used at Intercom to quickly turn around a group of teams that was in trouble.
The power of peer coaching for engineering managersOne of the most influential thing on your growth are the peers you work with. They provide you new perspectives, play devil advocate or brainstorm new ideas. Find them as engineering manager.
Software documentation in growing companiesFinding the right balance between moving fast and keeping high-quality documentation is very hard for growing engineering organisations. If you feel that these problems are present in your company, experiment with Architecture Design Records.
Learnings from Nonviolent CommunicationNonviolent Communication is a game changer tool for every people manager. It helps creating meaningful relationships by embracing any emotion, good or bad. And the best part is, it is even more useful in private life.