Notes on independent software, product decisions, interface craft, and what we learn while shipping.https://makerportal.ai/https://makerportal.ai/blog/2026-08-11-itria-hackernews-49245161/https://makerportal.ai/blog/2026-08-11-itria-hackernews-49245161/PrivateRedact's auto-selection limits Ollama inference to 4.0 tok/s on an 8 GB Pi 5 by defaulting to CPU memory, ignoring unified bandwidth.Tue, 11 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/auralinter-agentic-dsp-verification/https://makerportal.ai/blog/auralinter-agentic-dsp-verification/How RAG-grounded retrieval and a clang++ verifier catch a hallucinated biquad formula a single LLM call ships — and why AuraLinter isn't fully on-device.Fri, 17 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/auralinter-arxiv-260825404v1/https://makerportal.ai/blog/auralinter-arxiv-260825404v1/CSAVocoder renders binaural and FOA audio from mel-spectrograms and a 7D pose stream.Thu, 27 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/auralinter-arxiv-260904975v1/https://makerportal.ai/blog/auralinter-arxiv-260904975v1/The module named Spatial Edit supplies 91% of SwanWeave's fidelity win over SmartDJ but only 5% of the spatial one — pretraining does the spatial work.Mon, 07 Sep 2026 00:00:00 GMThttps://makerportal.ai/blog/binaural-head-headphone-measurement/https://makerportal.ai/blog/binaural-head-headphone-measurement/Class-compliant multichannel USB audio capture, a loaded 1 V / 103 dB SPL nominal reference at 1 kHz, and binaural recording analyzed on iPhone.Sun, 09 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/biquadia-github-1095330682/https://makerportal.ai/blog/biquadia-github-1095330682/Two figures in LLMKube's GKE benchmark are wrong: prompt processing is 35.4×, not 66×, and token generation is 13.9×, not 17×.Sun, 09 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/biquadia-github-1138007869/https://makerportal.ai/blog/biquadia-github-1138007869/Wax's own benchmark puts hybrid recall p95 at 6.1 ms — six times above the "sub-millisecond" label — and the figure omits which index engine ran.Sat, 22 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/biquadia-github-1171688936/https://makerportal.ai/blog/biquadia-github-1171688936/Driving Apple's Neural Engine through private frameworks: 172.4 tok/s on ANE vs 283 on CPU, and delta compilation that pays for itself in 1.21 steps.Wed, 05 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/biquadia-github-1173023018/https://makerportal.ai/blog/biquadia-github-1173023018/Espresso's 3.41× ANE speedup over CoreML, read against a 70 µs dispatch floor, sets a 168.7 µs lower bound on CoreML per-token overhead.Fri, 14 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/biquadia-github-1233899535/https://makerportal.ai/blog/biquadia-github-1233899535/Bandwidth calculation for Presspeech's Parakeet model on M4 Max shows inference takes ~1% of latency, with pipeline dispatch consuming the rest.Wed, 05 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/biquadia-github-1267734215/https://makerportal.ai/blog/biquadia-github-1267734215/ANEForge reaches 16.8 tok/s on Qwen3-8B (up from 7.4), hitting the verify(K)≈verify(1) ceiling at 29.6 GB/s vs Orion's 42.8 GB/s floor.Mon, 10 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/biquadia-github-1349261167/https://makerportal.ai/blog/biquadia-github-1349261167/mac-tool-kit's 7.8°C hysteresis band splits 8.75:1, placing the fan stop at 27.0°C — a temperature inference or DSP sessions never return to mid-session.Sat, 29 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/biquadia-hackernews-49259339/https://makerportal.ai/blog/biquadia-hackernews-49259339/M4 Max llama.cpp inference consumes 99.0–99.8% of its 400 GB/s bandwidth, leaving no slack to absorb VM overhead — a direct constraint on near-native…Wed, 12 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/blexar-ble-uart-bridge/https://makerportal.ai/blog/blexar-ble-uart-bridge/Why HM-10's single-characteristic UART and the Nordic UART Service aren't interchangeable, and the newline-framing gotcha that only appears at low MTU.Fri, 17 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/designing-the-one-handed-moment/https://makerportal.ai/blog/designing-the-one-handed-moment/A field note on glanceability, reach, and building interfaces for the imperfect moments where software is actually used.Sat, 04 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/elevenlabs-web-audio-streaming-latency/https://makerportal.ai/blog/elevenlabs-web-audio-streaming-latency/Browser latency ledger for ElevenLabs TTS, Web Audio scheduling, DSP, and output—plus mistakes that make a fast demo sound broken.Sun, 19 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/elevenlabs-webassembly-audioworklet-streaming/https://makerportal.ai/blog/elevenlabs-webassembly-audioworklet-streaming/How to build sub-100ms real-time voice streaming applications in the browser using ElevenLabs API, WebAssembly decoders, and AudioWorklet ring buffers.Wed, 29 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/field-note-hf-tencent-hy4-preview/https://makerportal.ai/blog/field-note-hf-tencent-hy4-preview/Tencent's Hy4 ships Apache 2.0 weights with 49B activated parameters and 1M-token context, but four unexplained residual streams block anything beyond…Fri, 28 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/inside-biquadia/https://makerportal.ai/blog/inside-biquadia/How privacy, real-time constraints, and the character of physical sound shaped MakerPortal’s newest release.Sun, 28 Jun 2026 00:00:00 GMThttps://makerportal.ai/blog/iphone-18-pro-vs-iphone-17-pro-launch-iphone-18-pro/https://makerportal.ai/blog/iphone-18-pro-vs-iphone-17-pro-launch-iphone-18-pro/Apple's own spec pages for the iPhone 18 Pro and iPhone 17 Pro, diffed: 16 sections changed, 15 identical. Every figure copied from the two documents.Fri, 11 Sep 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-arxiv-260902789v1/https://makerportal.ai/blog/itria-arxiv-260902789v1/Clang LibTooling pipeline extracts AST and CFG features from C++ LLM inference code. In-sample macro F1 is 0.40; leave-one-project-out drops to 0.112.Thu, 03 Sep 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-github-1051796718/https://makerportal.ai/blog/itria-github-1051796718/ETOS-LLM-Studio pins watchOS inference to the CPU path, so bandwidth divided by model size sets the ceiling — 9.4 tok/s on a Pi 5 at 17 GB/s.Sat, 08 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-github-1069760430/https://makerportal.ai/blog/itria-github-1069760430/OpenMed's 24–33× MLX speedup is a compute advantage, not a memory one. Batch throughput (3.3× CPU, 2.2× MLX) shows both backends are partially…Wed, 09 Sep 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-github-1164344011/https://makerportal.ai/blog/itria-github-1164344011/imp reports a 42-48% decode gap over llama.cpp on consumer Blackwell. The 5090's own 5090D comparison shows the gap is just a bandwidth model.Wed, 05 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-github-725205304/https://makerportal.ai/blog/itria-github-725205304/Bandwidth-bound decode math on RTX 5090D measurements backs out that only ~32% of Qwen3.6-35B's nominal weights are active per forward pass.Fri, 07 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-github-867433720/https://makerportal.ai/blog/itria-github-867433720/Two 7B Q4 models exceed an 8 GB ceiling when loaded concurrently. Serial residency via llama-swap reclaims 4.2 GB of KV cache headroom.Sat, 15 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-hackernews-49321813/https://makerportal.ai/blog/itria-hackernews-49321813/Pruning DeepSeek-V4-Flash-0731 cuts safetensors by 32.63%, but each 0.1-unit code NLL gain costs 0.245 units of WikiText degradation.Mon, 17 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-hackernews-49511882/https://makerportal.ai/blog/itria-hackernews-49511882/A llama.cpp fork adds block-granular KV cache streaming to the CUDA server path using pinned host memory and a bounded pool, enabling 262K context on a…Tue, 01 Sep 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-hackernews-49523381/https://makerportal.ai/blog/itria-hackernews-49523381/The same gpt-oss-20b weights produce scores from ~0% to ~87% depending on template, tool format, wire API, and reasoning effort.Wed, 02 Sep 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-hf-davidau-qwen38-27b-cold-fusion-gain-v11-nm-dau-neo-max-mtp-gguf/https://makerportal.ai/blog/itria-hf-davidau-qwen38-27b-cold-fusion-gain-v11-nm-dau-neo-max-mtp-gguf/GGUFs of a Qwen3-27B fine-tune with 75% linear attention layers and ~2/3 thinking-token reduction — benchmark gaps and unverified quality claims noted.Sun, 23 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-hf-hauhaucs-qwen38-27b-uncensored-hauhaucs-aggressive-mtp-gguf/https://makerportal.ai/blog/itria-hf-hauhaucs-qwen38-27b-uncensored-hauhaucs-aggressive-mtp-gguf/A 3.02× speedup exceeds the 3.00× ceiling for depth-2 speculative decoding, placing FastMTP's draft depth at 3 or more with an implied acceptance rate of…Wed, 19 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/itria-hf-peculiar-ragdoll-tiel-coder-35b-a3b-gguf/https://makerportal.ai/blog/itria-hf-peculiar-ragdoll-tiel-coder-35b-a3b-gguf/Tiel-Coder's README places benchmark figures side by side across two quantization regimes, two chat templates, and unstated precision.Mon, 31 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/lattepanda-vs-jetson-orin-edge-ai-benchmarks/https://makerportal.ai/blog/lattepanda-vs-jetson-orin-edge-ai-benchmarks/Hardware benchmark comparing x86 SBCs, ARM CUDA Tensor Cores, and USB Edge TPUs for local vision, speech, and embedding inference.Tue, 28 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/litefs-multi-region-sqlite/https://makerportal.ai/blog/litefs-multi-region-sqlite/Mental model for LiteFS multi-region SQLite: single-primary writes, local replicas, fly-replay forwarding, and honest latency modeling.Sun, 19 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/local-decode-bench-2026-08-26/https://makerportal.ai/blog/local-decode-bench-2026-08-26/Decode, prefill and first-token measured on Apple M3 Ultra, 256 GB: 9.8–139 tok/s across 9 models, median of 3 runs, prompt cache defeated.Wed, 26 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/motionlink-arxiv-260816722v1/https://makerportal.ai/blog/motionlink-arxiv-260816722v1/BEM-synthesised HRTFs match measured HRTFs on every polar localisation metric (N = 20), but torso-omission error at low rear elevations is untested.Tue, 18 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/motionlink-headphone-motion-api/https://makerportal.ai/blog/motionlink-headphone-motion-api/Quaternion attitude from the Headphone Motion API, why the reference frame resets on every launch, and the recentering pattern that fixes it.Wed, 15 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/notiary-hackernews-49362669/https://makerportal.ai/blog/notiary-hackernews-49362669/Minna indexes documents into a hybrid vector-plus-full-text store with citation-grounded chat, but requires macOS 26 beta and ships PostHog telemetry on…Thu, 20 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/notiary-on-device-semantic-search/https://makerportal.ai/blog/notiary-on-device-semantic-search/Notiary's semantic brain: converting a sentence-transformer to CoreML, deterministic storage math, and the fixed-length-input problem long notes hit.Fri, 17 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/nymic-arxiv-260827360v1/https://makerportal.ai/blog/nymic-arxiv-260827360v1/XTTSv2 reaches 0.49 EER on CommonVoice via speaker-embedding swaps, bounded by one ECAPA2 probe and seven of sixteen supported languages.Sun, 30 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/nymic-knn-voice-conversion/https://makerportal.ai/blog/nymic-knn-voice-conversion/How kNN-VC voice conversion works: WavLM feature extraction, nearest-neighbor matching, HiFiGAN vocoding, and voice-bank coverage limits.Fri, 17 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/on-device-ai-arxiv-260822110v1/https://makerportal.ai/blog/on-device-ai-arxiv-260822110v1/A 64-shape primitive matrix and ANE counters reveal routing follows operation form, with a ~0.77 bytes-per-token law holding across fp16, int8, and 2-bit…Tue, 25 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/on-device-ai-arxiv-260903363v1/https://makerportal.ai/blog/on-device-ai-arxiv-260903363v1/The 30% figure in the PACodec paper comes from two separate comparisons at different sample rates, baseline bitrates, and dataset sizes.Fri, 04 Sep 2026 00:00:00 GMThttps://makerportal.ai/blog/on-device-ai-github-1116516142/https://makerportal.ai/blog/on-device-ai-github-1116516142/Fun-ASR-Nano's 16x, 3–5x, and ~50% speedup figures are attached to different—or missing—conditions that the README never aligns.Mon, 24 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/on-device-ai-github-286898814/https://makerportal.ai/blog/on-device-ai-github-286898814/The CoreML-Models repo lists file sizes that are directly comparable across models, but the speed and parameter claims embedded in the same README…Thu, 10 Sep 2026 00:00:00 GMThttps://makerportal.ai/blog/on-device-ai-github-614149835/https://makerportal.ai/blog/on-device-ai-github-614149835/AutoSubs' model table uses one star scale across models with different corpora, quantization levels, and language sets.Sun, 06 Sep 2026 00:00:00 GMThttps://makerportal.ai/blog/on-device-ai-hackernews-49703818/https://makerportal.ai/blog/on-device-ai-hackernews-49703818/A 3.5-bit GGUF quantization of the 180B Qwen3.8-Flash-Next model runs on 2× RTX 3090s and scores 2.85 points above its BF16 reference on MMLU-Pro.Tue, 15 Sep 2026 00:00:00 GMThttps://makerportal.ai/blog/on-device-ai-hf-obliteratus-qwen38-27b-obliterated/https://makerportal.ai/blog/on-device-ai-hf-obliteratus-qwen38-27b-obliterated/V1's −6.0pp and V2's −0.3pp MMLU gaps come from sample sizes a factor of ten apart; V1's arithmetic doesn't reproduce — the figures are not…Fri, 21 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/on-device-vector-search-swift/https://makerportal.ai/blog/on-device-vector-search-swift/How to build distraction-free local semantic search on iOS using CoreML feature extraction and native Swift HNSW index graphs, grounded in Notiary.Mon, 03 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/running-3b-llms-on-microcontrollers/https://makerportal.ai/blog/running-3b-llms-on-microcontrollers/Why parameter count isn't the bottleneck for on-device LLMs: memory bandwidth, GGUF/INT4 quantization math, and ANE lessons from Thumbdash and itria.Mon, 03 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/thumb-dash-github-1192318297/https://makerportal.ai/blog/thumb-dash-github-1192318297/The step-count formula in buun-llama-cpp's examples reduces to exactly 10 degradation steps per attention layer — 6 codec tiers × 2 KV sides minus 2.Sun, 16 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/thumb-dash-hackernews-49279500/https://makerportal.ai/blog/thumb-dash-hackernews-49279500/Measured Q4 GGUF sizes show two 7B models consume 8.4 GB, 0.4 GB over DLLM's ceiling, leaving no room for the embed model or KV-cache.Thu, 13 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/thumbdash-hackernews-49434442/https://makerportal.ai/blog/thumbdash-hackernews-49434442/Document tasks yield acceptance 0.615 vs. 0.300 for reasoning — a 2.05× gap quant level and hardware cannot explain; both run Q8KP on RTX 6000 Ada.Wed, 26 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/uma-8-beamforming-iphone/https://makerportal.ai/blog/uma-8-beamforming-iphone/Real-time SRP-PHAT direction finding and MVDR superdirective beamforming with the miniDSP UMA-8 mic array on an iPhone in Biquadia.Sun, 09 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/webaudio-audioworklet-dsp-voice-ai/https://makerportal.ai/blog/webaudio-audioworklet-dsp-voice-ai/How to build zero-glitch AudioWorklet streaming, Schroeder reverb tails, and formant filtering for real-time speech and Voice AI in the browser.Mon, 03 Aug 2026 00:00:00 GMThttps://makerportal.ai/blog/webgpu-benchmark-browser/https://makerportal.ai/blog/webgpu-benchmark-browser/Verification-first WebGPU benchmark methodology: warm-up, queue completion, CPU cross-checks, honest FLOP accounting, and cloud crossover math.Sun, 19 Jul 2026 00:00:00 GMThttps://makerportal.ai/blog/why-we-ship-small/https://makerportal.ai/blog/why-we-ship-small/The case for narrow products, strong opinions, and releases that can still surprise their makers.Sat, 11 Jul 2026 00:00:00 GMT