Notes from building real systems.

Each post explains one system in plain English: what it does, how it works, and what went wrong when I built it.

Whisper Architecture: From Audio Waveform to Text

Follow Whisper from the raw waveform to the final text: the encoder, the decoder, the self-KV and cross-KV caches, and the decoding loop.

Learn the main terms of a secure enclave system, such as keys, hashes, encryption, and attestation, and the question that each term answers.

FFmpeg is a set of libraries and tools that read, change, and write audio and video. Learn its parts and how they work together.

Intel QAT hardware does TLS crypto for the CPU. This post shows how to build, patch, and use it, with a 56.6x speedup.

We found and fixed three bugs in a computer vision benchmark system: too much memory use, full disks, and tasks that ran again and again.

Run a large language model on RunPod GPU cloud. Use the REST API, and fix the disk size and the CUDA library conflict.

Learn how homomorphic encryption lets a computer do math on encrypted data, and how the SEAL library uses it with BFV and CKKS.

AMX is hardware in Intel CPUs that multiplies blocks of numbers at once, so you can run AI faster without a graphics card.

RAG chunks lose their source context. Add a VLLM-written description to each chunk before you embed it.

Mastering asynchronous task processing with Celery, Redis, and RabbitMQ for high-throughput distributed systems.

Deploying and scaling containerized applications on AWS with ECS Fargate for zero-infrastructure management.