MakerPortal / field notes
Working ideas about independent software, product judgment, interface craft, and the useful tension between speed and care.
Newsletter · Privacy-first
Privacy-first newsletter via Buttondown — no tracking pixels. Get new notes + archive drops first. Free snippet from BLExAR as lead magnet.
No tracking, unsubscribe anytime · RSS at /rss.xml · Privacy-first per D-014
Showing all 57 field notes
01
Readout / On-Device AI
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.
By Joshua Hrisko · 5 min read
Sep 15, 2026
02
Launch diff
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.
By Joshua Hrisko · 11 min read
Sep 11, 2026
03
Comparison / On-Device AI
The CoreML-Models repo lists file sizes that are directly comparable across models, but the speed and parameter claims embedded in the same README…
By Joshua Hrisko · 6 min read
Sep 10, 2026
04
Correction / On-Device AI
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…
By Joshua Hrisko · 5 min read
Sep 09, 2026
05
Correction / Audio DSP
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.
By Joshua Hrisko · 5 min read
Sep 07, 2026
06
Comparison / On-Device AI
AutoSubs' model table uses one star scale across models with different corpora, quantization levels, and language sets.
By Joshua Hrisko · 8 min read
Sep 06, 2026
07
Comparison / On-Device AI
The 30% figure in the PACodec paper comes from two separate comparisons at different sample rates, baseline bitrates, and dataset sizes.
By Joshua Hrisko · 6 min read
Sep 04, 2026
08
Paper / Local LLMs
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.
By Joshua Hrisko · 6 min read
Sep 03, 2026
09
Correction / Local LLMs
The same gpt-oss-20b weights produce scores from ~0% to ~87% depending on template, tool format, wire API, and reasoning effort.
By Joshua Hrisko · 5 min read
Sep 02, 2026
10
Readout / Local LLMs
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…
By Joshua Hrisko · 6 min read
Sep 01, 2026
11
Comparison / On-Device AI
Tiel-Coder's README places benchmark figures side by side across two quantization regimes, two chat templates, and unstated precision.
By Joshua Hrisko · 5 min read
Aug 31, 2026
12
Paper / Audio DSP
XTTSv2 reaches 0.49 EER on CommonVoice via speaker-embedding swaps, bounded by one ECAPA2 probe and seven of sixteen supported languages.
By Joshua Hrisko · 4 min read
Aug 30, 2026
13
Correction / Metal & the Neural Engine
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.
By Joshua Hrisko · 6 min read
Aug 29, 2026
14
Readout / Trend-grounded
Tencent's Hy4 ships Apache 2.0 weights with 49B activated parameters and 1M-token context, but four unexplained residual streams block anything beyond…
By Joshua Hrisko · 6 min read
Aug 28, 2026
15
Readout / Audio DSP
CSAVocoder renders binaural and FOA audio from mel-spectrograms and a 7D pose stream.
By Joshua Hrisko · 6 min read
Aug 27, 2026
16
Measurement / On-Device AI
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.
By Joshua Hrisko · 6 min read
Aug 26, 2026
17
Correction / Local LLMs
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.
By Joshua Hrisko · 6 min read
Aug 26, 2026
18
Paper / On-Device AI
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…
By Joshua Hrisko · 6 min read
Aug 25, 2026
19
Comparison / On-Device AI
Fun-ASR-Nano's 16x, 3–5x, and ~50% speedup figures are attached to different—or missing—conditions that the README never aligns.
By Joshua Hrisko · 5 min read
Aug 24, 2026
20
Readout / On-Device AI
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.
By Joshua Hrisko · 5 min read
Aug 23, 2026
21
Correction / On-Device AI
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.
By Joshua Hrisko · 6 min read
Aug 22, 2026
22
Comparison / On-Device AI
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…
By Joshua Hrisko · 6 min read
Aug 21, 2026
23
Readout / On-Device AI
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…
By Joshua Hrisko · 6 min read
Aug 20, 2026
24
Correction / On-Device AI
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…
By Joshua Hrisko · 6 min read
Aug 19, 2026
25
Paper / Audio DSP
BEM-synthesised HRTFs match measured HRTFs on every polar localisation metric (N = 20), but torso-omission error at low rear elevations is untested.
By Joshua Hrisko · 6 min read
Aug 18, 2026
26
Field note / Local LLMs
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.
By Joshua Hrisko · 6 min read
Aug 17, 2026
27
Field note / Local LLMs
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.
By Joshua Hrisko · 6 min read
Aug 16, 2026
28
Field note / Local LLMs
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.
By Joshua Hrisko · 6 min read
Aug 15, 2026
29
Field note / On-Device AI
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.
By Joshua Hrisko · 6 min read
Aug 14, 2026
30
Field note / Local LLMs
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.
By Joshua Hrisko · 5 min read
Aug 13, 2026
31
Field note / Metal & the Neural Engine
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…
By Joshua Hrisko · 6 min read
Aug 12, 2026
32
Field note / Local LLMs
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.
By Joshua Hrisko · 5 min read
Aug 11, 2026
33
Field note / On-Device AI
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.
By Joshua Hrisko · 7 min read
Aug 10, 2026
34
Hardware build / 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.
By Joshua Hrisko · 11 min read
Aug 09, 2026
35
Field note / Metal & the Neural Engine
Two figures in LLMKube's GKE benchmark are wrong: prompt processing is 35.4×, not 66×, and token generation is 13.9×, not 17×.
By Joshua Hrisko · 6 min read
Aug 09, 2026
36
Hardware build / Mic arrays
Real-time SRP-PHAT direction finding and MVDR superdirective beamforming with the miniDSP UMA-8 mic array on an iPhone in Biquadia.
By Joshua Hrisko · 10 min read
Aug 09, 2026
37
Field note / Local LLMs
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.
By Joshua Hrisko · 7 min read
Aug 08, 2026
38
Field note / Local LLMs
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.
By Joshua Hrisko · 7 min read
Aug 07, 2026
39
Field note / On-Device AI
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.
By Joshua Hrisko · 7 min read
Aug 05, 2026
40
Field note / On-Device AI
Bandwidth calculation for Presspeech's Parakeet model on M4 Max shows inference takes ~1% of latency, with pipeline dispatch consuming the rest.
By Joshua Hrisko · 7 min read
Aug 05, 2026
41
Field note / Edge Hardware Systems
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.
By Joshua Hrisko · 7 min read
Aug 05, 2026
42
Field note / On-Device Vector Search
How to build distraction-free local semantic search on iOS using CoreML feature extraction and native Swift HNSW index graphs, grounded in Notiary.
By Joshua Hrisko · 5 min read
Aug 03, 2026
43
Field note / Edge Hardware Systems
Why parameter count isn't the bottleneck for on-device LLMs: memory bandwidth, GGUF/INT4 quantization math, and ANE lessons from Thumbdash and itria.
By Joshua Hrisko · 6 min read
Aug 03, 2026
44
Field note / Web Audio & Neural DSP
How to build zero-glitch AudioWorklet streaming, Schroeder reverb tails, and formant filtering for real-time speech and Voice AI in the browser.
By Joshua Hrisko · 5 min read
Aug 03, 2026
45
Field note / Web Audio & Synthetic Voice
How to build sub-100ms real-time voice streaming applications in the browser using ElevenLabs API, WebAssembly decoders, and AudioWorklet ring buffers.
By Joshua Hrisko · 8 min read
Jul 29, 2026
46
Field note / Edge Hardware Systems
Hardware benchmark comparing x86 SBCs, ARM CUDA Tensor Cores, and USB Edge TPUs for local vision, speech, and embedding inference.
By Joshua Hrisko · 9 min read
Jul 28, 2026
47
Field note / Voice infrastructure
Browser latency ledger for ElevenLabs TTS, Web Audio scheduling, DSP, and output—plus mistakes that make a fast demo sound broken.
By Joshua Hrisko · 7 min read
Jul 19, 2026
48
Field note / Edge databases
Mental model for LiteFS multi-region SQLite: single-primary writes, local replicas, fly-replay forwarding, and honest latency modeling.
By Joshua Hrisko · 8 min read
Jul 19, 2026
49
Field note / GPU deployment
Verification-first WebGPU benchmark methodology: warm-up, queue completion, CPU cross-checks, honest FLOP accounting, and cloud crossover math.
By Joshua Hrisko · 8 min read
Jul 19, 2026
50
Field note / AuraLinter
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.
By Joshua Hrisko · 5 min read
Jul 17, 2026
51
Field note / BLExAR
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.
By Joshua Hrisko · 5 min read
Jul 17, 2026
52
Field note / Notiary
Notiary's semantic brain: converting a sentence-transformer to CoreML, deterministic storage math, and the fixed-length-input problem long notes hit.
By Joshua Hrisko · 4 min read
Jul 17, 2026
53
Field note / nymic
How kNN-VC voice conversion works: WavLM feature extraction, nearest-neighbor matching, HiFiGAN vocoding, and voice-bank coverage limits.
By Joshua Hrisko · 4 min read
Jul 17, 2026
54
Field note / MotionLink
Quaternion attitude from the Headphone Motion API, why the reference frame resets on every launch, and the recentering pattern that fixes it.
By Joshua Hrisko · 5 min read
Jul 15, 2026
55
Studio philosophy
The case for narrow products, strong opinions, and releases that can still surprise their makers.
By Joshua Hrisko · 3 min read
Jul 11, 2026
56
Interface notes
A field note on glanceability, reach, and building interfaces for the imperfect moments where software is actually used.
By Joshua Hrisko · 4 min read
Jul 04, 2026
57
Build log / 004
How privacy, real-time constraints, and the character of physical sound shaped MakerPortal’s newest release.
By Joshua Hrisko · 5 min read
Jun 28, 2026
No field note carries that tag yet.
