Ever wonder why your MILP model slows to a crawl as soon as you add a few binary variables? That’s because every binary variable doubles the number of possible solutions the solver has to explore. Add 10 binaries → 1,024 combinations. Add 30 binaries → over a billion. This is why “scalability” in optimization isn’t about how fast your laptop is, it’s about how you formulate the problem. A few tricks I’ve learned over the years: ⚡ Replace loose Big-M constraints with indicator constraints or tighter formulations. ⚡ Use network flow structures instead of generic binaries (solvers love them). ⚡ Aggregate where you can, disaggregate only where it matters. I’ve seen models go from hours to seconds just by re-thinking the formulation. Most of the time, it’s not about clever heuristics, but rather respecting the math and crafting the right formulation. 💬 Curious… what’s the biggest “aha” moment you’ve had in making an optimization model scale?

Finding the right electronic components for a circuit board is harder than it should be. Engineers need to balance strict requirements, ensure compatibility across the design, and dig through datasheets, often spending hours just to identify a single part. Multiply that by hundreds of parts in a design, and the process becomes a massive bottleneck. That’s why I’ve been building a way to automate this entire workflow using a swarm of AI agents. With just one prompt, you can now discover and validate the parts you need, without the manual grind. I just opened the beta -> https://solderable.dev If you’ve ever been frustrated by sourcing parts, I’d love for you to try it out and share your feedback.

Five reasons you might want to consider a local model: 1. Local models don't leak your information 2. Local models run faster 3. Local models don't get discontinued by the vendor 4. Local models can be arbitrarily fine-tuned/modified 5. Local models don't change without notice There are many reasons as well for staying away from local models. Personally, I always recommend that companies start experimenting with online models (gpt, gemini, claude) because they tend to be better and easier to use. Once you determine you can solve a problem with a proprietary model, then you can decide whether using a local model is the right next step for you.

8px base unit for all spacing creates visual consistency automatically. Use multiples of 8 for margins, padding, and element spacing: 8px, 16px, 24px, 32px. A startup's website felt "messy" even though individual elements looked good. The problem: random spacing values throughout the design. We standardized on 8px increments. The site immediately felt more cohesive and professional. Mathematical spacing creates unconscious harmony.

Anamorphic lenses often have a 'waterfall' distortion effect, where vertical lines at the edges of the frame seem to curve inwards. It's a type of barrel distortion that's unique to their design. Some people hate it and try to correct it in software. Others lean into it as part of the lens's character. There's no right or wrong, but it's good to know it's a feature, not a flaw with your specific lens.

Discover how intelligent search and AI-driven predictive commands transform the PCB design experience! In this video, David Haboud showcases powerful features that streamline your workflow and boost productivity, whether you're new to PCB design or a seasoned professional. What you'll learn: 🔍 How global search eliminates the need to memorize complex commands ⌨️ Finding functions using flexible keyword searches that understand industry terminology 🧠 Leveraging AI-driven predictive commands that adapt to your personal workflow 📈 Accessing built-in machine learning that anticipates your next design moves 📖 Using integrated documentation search without leaving your design environment See how these intelligent features bridge skill gaps in your team and help everyone work more efficiently. With AI-powered assistance throughout the interface, you'll spend less time hunting for commands and more time creating exceptional designs.

Discover how intelligent search and AI-driven predictive commands transform the PCB design experience! In this video, David Haboud showcases powerful features that streamline your workflow and boost productivity, whether you're new to PCB design or a seasoned professional. What you'll learn: 🔍 How global search eliminates the need to memorize complex commands ⌨️ Finding functions using flexible keyword searches that understand industry terminology 🧠 Leveraging AI-driven predictive commands that adapt to your personal workflow 📈 Accessing built-in machine learning that anticipates your next design moves 📖 Using integrated documentation search without leaving your design environment See how these intelligent features bridge skill gaps in your team and help everyone work more efficiently. With AI-powered assistance throughout the interface, you'll spend less time hunting for commands and more time creating exceptional designs.

While exploring more about system design I came across another classic rate limiting algorithm: The Fixed Window Counter Algorithm. Here’s how it works. The timeline is divided into fixed-sized time windows, and each window has a counter. Every request increments the counter by one. Once the counter hits the threshold, new requests are dropped until the next window begins. For example (refer image 1) , imagine a system that allows a maximum of 3 requests per second. If more than 3 requests arrive in that second, the extra ones are dropped. Sounds simple, right? And it is. The algorithm is easy to understand, memory efficient, and resetting quotas at the end of a window works well in certain use cases. But there’s a catch (refer image 2) . A burst of requests at the edges of two windows can let more traffic through than intended. Say a system allows 5 requests per minute. If 5 requests arrive at the end of one minute and another 5 come in immediately at the start of the next, that’s 10 requests within a 60-second span. Twice the allowed limit. This is the major drawback of the Fixed Window Counter Algorithm. This is a reminder that even simple algorithms come with trade-offs. To address this limitation, the Sliding Window Log Algorithm offers a more adaptive solution, which I’ll explore in my next post. #SystemDesign #Algorithm

Reduction. Not minimalism. Precision... The best creative ideas don’t always add. Sometimes, they subtract. And in doing so, they hit harder!!! When Larry and Sergey showed the Google homepage to investors, it was just a white screen. One search box. Nothing else. The room went quiet. They waited for more. But that was it. That was the product. And that moment stuck—not because it dazzled, but because it didn’t try to. It trusted the clarity of intent. Apple’s “Think Different” never explained the tech. Just two words. And a feeling. Nike’s “Just Do It” didn’t need context. Three words. And a movement. Pixar’s Up opened with silence. No dialogue. Just emotion. That sequence said more than most scripts ever could. Even the grand temple at Ellora wasn’t built. It was chiseled out of a single rock - Building downwards! No scaffolding. No additions. Just subtraction with purpose. The grand temple emerged from what was removed, not what was added... I’ve seen this outside work too. There are moments where I start with too much—too many words, too many layers—and then pare it down for my daughter. Sometimes, just one sentence. The truth. The fact. And suddenly, it lands. Not because it’s clever. But because it’s clear. That’s what reduction really is. Not emptiness. Not absence. But the discipline to hold back... To trust the audience. To let the idea breathe. Because when you strip away the noise, what remains isn’t less—it’s essential. It’s the part that stays and builds over time! So here’s a thought— What’s one idea you made stronger by doing less? Drop it below. Let’s build a thread of subtraction.

Qwen3-Next just landed in Hugging Face Transformers and the config tells a story. This isn’t just “Qwen scaled up.” The architecture is fundamentally hybrid and built for long-context, high-throughput inference at scale. A few things stand out: Hybrid attention schedule: 3 layers of linear attention followed by 1 full softmax attention, repeated across 48 layers. Optimized for long sequence efficiency without giving up global context. Context length: 32,768 tokens. Rotary embeddings applied to only 25% of Q/K dimensions (partial RoPE), stabilizing training while keeping costs down. The config also exposes rope_scaling (linear, dynamic, YaRN, LongRoPE, LLaMA3), making context extension first-class. Attention efficiency: 16 query heads, but only 2 key/value heads. That’s an 8:1 GQA ratio, collapsing KV cache size by ~8×. MoE capacity: 512 experts, with top-10 routing per token (most MoEs use top-2), plus a shared expert pathway. This massively increases capacity while keeping inference sparse. Other specs: • 48 decoder layers, hidden size 2048, MLP dimension 5632 • Vocabulary size 151,936 (clearly multilingual + code-ready) • RMSNorm, SiLU activation • Predefined tensor and pipeline parallel partitioning baked directly into the config Takeaway: Qwen3-Next is a hybrid design with linear + full attention, partial RoPE, extreme KV compression, and high-sparsity MoE with top-10 routing. It’s engineered for long context, throughput, and distributed training. The architecture alone makes a statement about where LLM design is going: efficiency and scale without sacrificing capacity.