If anyone is interested in doing this: https://dev.meta.ai/events/global-hackathon?utm_source=social-li&utm_medium=M4D&utm_campaign=organic&utm_content=aihackathon There is some money if you do happen to be a finalist.

Had my hiring-manager screen for Databricks Specialist Solutions Architect, Data Engineering & Warehousing, Financial Services. It was booked for an hour, but he finished his questions in about 30 minutes. That alone had me wondering whether he got the signal he needed — good or bad. The strongest part was walking through a cash-forecasting project I owned end to end: business discovery, Snowflake architecture, dbt, ML forecasting, and Tableau. That story landed. Where I got exposed was when he went one level deeper. I mentioned gradient boosting, and he asked me to explain it. I stumbled and started talking about train/validation/test splits instead of actually explaining boosting. Then I talked about using an LLM to analyze financial statements. He asked a simple but very good question: “If the math is deterministic, why use an LLM?” That one stuck with me. The better answer is: let code handle extraction and calculations, and let the LLM handle interpretation and judgment. I also rambled through an influence story about convincing leadership to move historical data from SQL Server to Snowflake. The story was fine, but he had to pull the actual decision out of me. My biggest takeaway: For a Solutions Architect interview, the high-level architecture story only gets you so far. You have to be ready for the next question. If you name a model, explain how it works. If you use an LLM, explain why it belongs there. If you influenced a decision, say what changed. That was probably the biggest lesson from this one: depth matters just as much as storytelling.

Hey @Alex Galea! I want to join your Wednesday calls. I just have a call that I have to attend from 11 am to 11.30 am. Any chance you plan to do alternative times for folks to attend? Just curious

Quick context: I'm the technical co-founder at an AI focus-group moderation startup (voice-based, live multi-participant — think a research platform, not a chatbot). This is actually my first hands-on build. I spent 15+ years as a Petroleum Engineer before doing a Master's in Applied AI, and while the program was great for the ML/AI side, it didn't cover much in the way of software engineering fundamentals or system design — so a lot of this is me learning architecture by doing, in real time, on a live product. I've been heads-down on the actual moderation engine (LiveKit + STT/TTS + an AI avatar) and just finished a client-facing frontend for my non-technical co-founders using Claude Design. Now I'm at the classic solo-founder fork: the frontend looks real, the backend (FastAPI + Postgres) is running but mostly unwired, and I'm trying to figure out the right order of operations before I just start bolting things together. Specifically wrestling with: 1. Auth-first vs. feature-first. Nothing downstream — user records, engagement data, analytics — means anything until real auth exists. Feels obvious in hindsight, but it means the "foundation" work is invisible to stakeholders who just want to see features ship. 2. What to actually track. I want to know where users drop off in a multi-step workflow (signup → contract → study setup → audience → instrument → session → report) without over-engineering a custom analytics stack before I even have real users. Self-hosted PostHog vs. just a Postgres events table vs. something else? 3. How much "systems design" to front-load. I'm working through Designing Data-Intensive Applications and the System Design Interview books to make better architecture calls now instead of paying for them later — but at 10 hrs/week solo, I can't read either cover to cover before shipping. Curious how others have picked their battles here (I landed on: data modeling + transactions + event/stream processing chapters now, replication/partitioning/consensus later).

Last night I built a cyberpunk coding game with Kimi K3 (prompt below) (sponsored video coming to my YouTube channel next week. this post is not sponsored) Here's what you should know about K3: - 2.8T parameter open weights model (largest ever) - 1 million token context window - Openly published architecture (kimi delta attention, attention residuals, etc..) The game ended up being pretty fun, so I wanted to share it. You can skip to the second half of the video to see me demo it. I used kimi code for this. Very capable agent harness and fun to use. I basically one-shotted the game. The secret was to include "Godot 4" in the prompt. That's the game engine I used. I've attached the full prompt if you want to check it out!