Quilio Engineering & Research Blog

Technical notes, DSP pipelines, systems architecture, and benchmark measurements from Quilio.

Articles

• Bun Rewrote in Rust. Zig Did Not Lose. (August 31, 2026, 8 min read): Bun's rewrite was an engineering decision inside one unusual runtime, not a verdict on Zig. The rivalry obscures the more useful lessons about ownership, testing, AI and project values.
• The Build Is Part of the Product (August 31, 2026, 5 min read): A product is not shippable when the feature works once. Builds, signing, CI and verification are part of the thing customers actually receive.
• Why Flow Treats Time as a Language Feature (August 31, 2026, 7 min read): Flow's future-looking idea is not new syntax for old programs. It makes evolution, effects, differentiation and deployment constraints part of the same executable model.
• CPU, GPU or Neural Engine? Choosing the Right Compute Unit for Inference (August 29, 2026, 6 min read): Short interactive workloads change the answer to where a model should run. We compare memory movement, dispatch overhead, latency and sustained performance.
• SmolLM2-135M at 155 Tokens per Second on Apple Silicon (August 28, 2026, 6 min read): A small C++ engine for on-device decode. Architecture, ARM NEON SIMD packing, int8 accuracy, and Snapdragon 8 Gen 3 thermal behaviour.
• How to Infer a Product's Architecture from Its Failure Modes (August 27, 2026, 6 min read): Observable failures expose the boundaries, queues and dependencies inside a product. A practical method for competitive analysis without source access.
• What Makes a Plugin Feel Expensive? (August 25, 2026, 5 min read): The perception of quality in an audio plugin comes from timing, control laws, metering and restraint as much as from its sound engine.
• Procedural Generation with Constraints, Not Randomness (August 23, 2026, 5 min read): A generated world becomes believable when its rules are declared, checked and allowed to produce consequences.
• Fixing JUCE Build Graph Duplication with schmoothie (August 21, 2026, 5 min read): How multi-target JUCE projects compile framework modules repeatedly, and how a clean cache layer reduces three-target builds to 0.40x CPU time.
• ABX Testing Without Fooling Yourself (August 20, 2026, 5 min read): A listening test is only useful when its procedure controls expectation, loudness, order and the limits of what the result can support.
• On-Device Push-to-Talk Dictation without Cloud Latency (August 15, 2026, 4 min read): Designing a zero-latency audio pipeline for Whisper models on macOS using local ring buffers, voice activity detection, and Core ML.
• Differentiable Bloch-Wave Electron Diffraction with faster-diffbloch (August 10, 2026, 5 min read): Accelerating dynamical electron scattering simulations with Metal GPU shaders and optimized CPU kernels, outperforming PyTorch by up to 1.80x.
• Time-Domain PSOLA with Independent Pitch and Formant Controls (July 30, 2026, 6 min read): Real-time pitch correction on a TD-PSOLA engine. F0 estimation, pitch mark allocation, Hann grain overlap-add, and vocal tract preservation.
• Measuring True Mac Memory Footprint for AI Agent Sessions (July 18, 2026, 5 min read): Why resident set size overcounts memory consumption by more than double, and how Vibe Meter reads physical footprint and compressed swap.
• Unified Pitch-Time Field: Audio Transforms at 9.8x Real Time (July 02, 2026, 5 min read): A continuous mathematical field for independent manipulation of harmonic pitch and temporal transients with morphable sample-accurate convolution.