Applied AI engineering. In-depth explorations of software engineering, deep learning, and the technologies that power modern systems.https://indepth.dev/https://indepth.dev/posts/2045/en/connecting-the-dots-where-hard-work-and-dreams-can-lead-you/https://indepth.dev/posts/2045/en/connecting-the-dots-where-hard-work-and-dreams-can-lead-you/How I went from being ready to quit tech to becoming a developer advocate, conference speaker, and founder of the largest Angular publication online — in two years of focused, relentless work.Tue, 08 Jan 2019 00:00:00 GMThttps://indepth.dev/posts/2025/en/floats-and-model-memory/https://indepth.dev/posts/2025/en/floats-and-model-memory/How IEEE-754 bit layouts decide what a model weighs in VRAM, why fp16 and bf16 are not the same format at half the size, and where quantization continues the same idea.Fri, 08 May 2026 00:00:00 GMThttps://indepth.dev/posts/2055/en/negative-sampling-contrastive-learning/https://indepth.dev/posts/2055/en/negative-sampling-contrastive-learning/word2vec introduced negative sampling to make a million-class softmax tractable, but the same trick — pull paired things together, push random things apart — turned out to underpin CLIP, sentence encoders, RAG retrievers, recommendation systems, and RLHF reward models. We walk through how negative sampling actually works, why it works, and where it shows up across modern ML.Fri, 29 May 2026 00:00:00 GMThttps://indepth.dev/posts/2043/en/word-embeddings-word2vec/https://indepth.dev/posts/2043/en/word-embeddings-word2vec/Before BERT, the breakthrough idea in NLP was that words could be represented as dense vectors where geometric relationships encode meaning. king − man + woman lands near queen. We walk through the distributional hypothesis, the inner workings of skip-gram and CBOW, explore the famous king–queen analogy, and see exactly where static word embeddings break — which is what motivates contextual models like BERT in the first place.Mon, 18 May 2026 00:00:00 GMThttps://indepth.dev/posts/1017/en/anatomy-of-a-developer-targeted-supply-chain-attack/https://indepth.dev/posts/1017/en/anatomy-of-a-developer-targeted-supply-chain-attack/A technical breakdown of a real malicious project sent to a developer — five attack vectors hidden behind a polished Web3 codebase, triggered by npm install or opening the folder in VSCode.Tue, 24 Mar 2026 00:00:00 GMThttps://indepth.dev/posts/1002/en/become-a-git-pro-by-learning-git-architecture/https://indepth.dev/posts/1002/en/become-a-git-pro-by-learning-git-architecture/A step-by-step practical exercises that explore git internalsThu, 10 Apr 2025 00:00:00 GMThttps://indepth.dev/posts/2029/en/binary-arithmetic-integers-and-floats/https://indepth.dev/posts/2029/en/binary-arithmetic-integers-and-floats/How computers actually add, subtract, and multiply numbers at the bit level — the clean modular arithmetic of two's complement and the exponent-aligning, mantissa-rounding dance of IEEE-754 floats.Tue, 20 Jan 2026 00:00:00 GMThttps://indepth.dev/posts/2013/en/building-model-armor-multi-layer-llm-safety-filtering/https://indepth.dev/posts/2013/en/building-model-armor-multi-layer-llm-safety-filtering/A hands-on guide to building a multi-layer LLM safety system inspired by Google Model Armor — combining rule-based filters, classifiers, LLM judges, prompt rewriting, and output moderation into a single pipeline.Wed, 29 Apr 2026 00:00:00 GMThttps://indepth.dev/posts/2030/en/context-assembly-strategies/https://indepth.dev/posts/2030/en/context-assembly-strategies/Every LLM agent makes invisible decisions about what to send the model each turn. We compare context-assembly strategies on a real bug-fix task, using pi — an agent framework that exposes those decisions as code.Tue, 05 May 2026 00:00:00 GMThttps://indepth.dev/posts/1005/en/decimal-binary-conversion-algorithms/https://indepth.dev/posts/1005/en/decimal-binary-conversion-algorithms/How to convert from binary to decimal and back, and the math that explains why the algorithms workSat, 20 Oct 2018 00:00:00 GMThttps://indepth.dev/posts/1003/en/dijkstra-was-right-recursion-should-not-be-difficult/https://indepth.dev/posts/1003/en/dijkstra-was-right-recursion-should-not-be-difficult/Recursion is one of the most important ideas in computer science. This article introduces topics and ideas that will help you when practicing recursive problems.Wed, 12 Nov 2025 00:00:00 GMThttps://indepth.dev/posts/1012/en/effect-of-hyperparameters-on-training/https://indepth.dev/posts/1012/en/effect-of-hyperparameters-on-training/Changing one thing at a time — learning rate, batch size, network depth, activation functions — and seeing exactly how each piece affects training.Thu, 02 Apr 2026 00:00:00 GMThttps://indepth.dev/posts/1007/en/exponent-bias-floating-point/https://indepth.dev/posts/1007/en/exponent-bias-floating-point/Understanding how offset binary (biased exponent) works, how it differs from two's complement, and why IEEE-754 uses it for floating-point exponentsTue, 25 Nov 2025 00:00:00 GMThttps://indepth.dev/posts/1023/en/gentle-introduction-into-compilers-lexical-analysis-scanner/https://indepth.dev/posts/1023/en/gentle-introduction-into-compilers-lexical-analysis-scanner/A guide to understanding ECMAScript (JavaScript) spec lexical grammar and TypeScript scannerThu, 05 Sep 2024 00:00:00 GMThttps://indepth.dev/posts/2006/en/gpu-overhead-why-mnist-trains-faster-on-cpu/https://indepth.dev/posts/2006/en/gpu-overhead-why-mnist-trains-faster-on-cpu/We profiled our MNIST training and found the GPU spends 99% of its time on overhead. Here is why small models can not saturate modern GPUs, and how we got a 45x speedup by reducing overhead instead of upgrading hardware.Sun, 05 Apr 2026 00:00:00 GMThttps://indepth.dev/posts/2026/en/how-floats-work-javascript-and-python/https://indepth.dev/posts/2026/en/how-floats-work-javascript-and-python/Same 64 bits, same quirks: why 0.1 + 0.2 isn't 0.3 in either language, where the safe-integer boundary comes from, and how language wrappers diverge on top of an identical float.Sat, 15 Feb 2025 00:00:00 GMThttps://indepth.dev/posts/1001/en/how-neural-networks-learn-backpropagation-gradient-descent/https://indepth.dev/posts/1001/en/how-neural-networks-learn-backpropagation-gradient-descent/A hands-on guide to understanding gradient descent and backpropagation — the core algorithms behind training neural networks.Wed, 18 Mar 2026 00:00:00 GMThttps://indepth.dev/posts/1020/en/how-trigger-dev-checkpoints-containers/https://indepth.dev/posts/1020/en/how-trigger-dev-checkpoints-containers/A deep dive into how Trigger.dev uses CRIU, Docker, and Buildah to snapshot running containers at await points — suspending compute billing and resuming execution hours later from the exact same memory state.Sat, 28 Mar 2026 00:00:00 GMThttps://indepth.dev/posts/1022/en/level-up-reverse-engineering-skills/https://indepth.dev/posts/1022/en/level-up-reverse-engineering-skills/Guidelines and principles from a developer who reverse-engineered modern web frameworksFri, 20 Jun 2025 00:00:00 GMThttps://indepth.dev/posts/1006/en/rounding-binary-numbers/https://indepth.dev/posts/1006/en/rounding-binary-numbers/Understanding the rules behind rounding binary fractions, explained with visual number line diagramsMon, 12 Nov 2018 00:00:00 GMThttps://indepth.dev/posts/2028/en/securing-a-web-app-you-dont-fully-understand/https://indepth.dev/posts/2028/en/securing-a-web-app-you-dont-fully-understand/A layered, practical playbook for hardening a web app whose internals you cannot or will not fully audit — covering secret scanning, edge filtering with a WAF, bot protection, dependency scanning with OSV-Scanner, SAST, and DAST.Mon, 27 Apr 2026 00:00:00 GMThttps://indepth.dev/posts/1004/en/signals-in-angular-deep-dive-for-busy-developers/https://indepth.dev/posts/1004/en/signals-in-angular-deep-dive-for-busy-developers/A comprehensive deep dive into Angular signals internals: reactive context, dependency graph, two-phase updates, change detection integration, and effectsTue, 06 Aug 2024 00:00:00 GMThttps://indepth.dev/posts/1008/en/training-and-experimenting-with-mnist/https://indepth.dev/posts/1008/en/training-and-experimenting-with-mnist/Putting backpropagation, gradient descent, and every concept from the previous article into practice — building and training a real neural network on handwritten digits.Thu, 26 Mar 2026 00:00:00 GMThttps://indepth.dev/posts/2007/en/understanding-cnns-convolutions-feature-maps-pooling/https://indepth.dev/posts/2007/en/understanding-cnns-convolutions-feature-maps-pooling/What convolutions actually compute, how feature maps capture patterns, and why pooling shrinks the representation — explained visually with interactive demos.Tue, 14 Apr 2026 00:00:00 GMT