# Human-friendly view → https://makerportal.ai/llms — raw below for LLMs, RAG, agentic browsers
# If you landed here as a human, the styled page explains everything. Bots: continue below (valid Markdown).

# MakerPortal — Independent AI & iOS Software Studio

> On-device first where possible, transparent where cloud is needed. San Francisco studio engineering iOS & iPadOS apps with Apple Silicon native neural processing (CoreML, Metal, ANE). Privacy posture per-app transparent.

MakerPortal ([makerportal.ai](https://makerportal.ai) / [makersportal.com](https://makersportal.com)) is an independent software studio established in 2023 in San Francisco, CA. Focus: focused utility, productivity, developer tools, audio, game, dashboard, and lifestyle applications — on-device first where possible (CoreML, Metal, ANE, ONNX, llama.cpp), transparent hybrid cloud for agentic DSP (AuraLinter backend with SSE, audio staged temporarily).

**Core principles:** On-device first where possible, transparent hybrid for agentic, no committee software, useful on day one, craft over clutter, one deliberate release at a time.

## Site map (hub) — empire IA (D-024)

Primary nav (6): Apps · Lab · Library · Blog · Shop · Studio

- [Home](https://makerportal.ai): Studio hub
- [Apps](https://makerportal.ai/apps): Catalog — 12 live, of which 11 are published on Apple's App Store and 1 is web/hybrid platform (product pages: 6 on *.makerportal.ai + 6 legacy bridge on makersportal.com/apps migrating)
- [Lab](https://makerportal.ai/lab): 49 browser research instruments
- [Console](https://makerportal.ai/console): Plain-language front end to the compute API — type a filter, room, coil or model/GPU pair and the registered solver runs client-side. Worked answers on the page are solved at build time.
- [Library](https://makerportal.ai/library): Daily engines + gear + tools hub
- [Edge AI Radar](https://makerportal.ai/library/edge-ai-radar): Daily GGUF/ONNX × board memory ceilings (deterministic)
- [Daily Signals](https://makerportal.ai/library/daily-signals): Daily trend backlog (under Library — not a top-level nav peer)
- [Field notes / blog](https://makerportal.ai/blog): Human-authored articles & craft
- [Shop](https://makerportal.ai/shop): Digital goods, code packs, archives
- [Studio](https://makerportal.ai/about) · [Team](https://makerportal.ai/team) · [Press](https://makerportal.ai/press) · [Advertise](https://makerportal.ai/advertise) · [Watch](https://makerportal.ai/watch) (scaffold)
- [Contact](https://makerportal.ai/contact) · [Privacy](https://makerportal.ai/privacy) · [Terms](https://makerportal.ai/terms)
- [RSS](https://makerportal.ai/rss.xml) · [Sitemap](https://makerportal.ai/sitemap.xml) · [Agent map](https://makerportal.ai/llms)

## Application Matrix — 12 Live (Featured indicated with *)

Full studio catalog available at [makerportal.ai/apps](https://makerportal.ai/apps). 11 of the 12 are published on Apple's App Store under the artist **Maker Portal LLC** and install directly from the link on each line; the remaining 1 is web/hybrid platform. Each product is also canonical on its own subdomain or legacy journal host:

**App Store prices below are as Apple reported them on 18 September 2026 and are indicative only** — the price shown on the App Store at the time of purchase applies.

- [*Articulance — On-device executive presence coach](https://articulance.makerportal.ai): Private executive presence coach that runs entirely on-device: rehearse out loud and get timestamped findings with audio receipts, live pace coaching, a teleprompter that listens, and The Room — mock interviews with five neural interviewers. One-time purchase, no account, no cloud. (iOS (iPhone + iPad)). Category: Productivity. Tech: On-device Speech, Neural Interviewers, Live Pace Coaching. App Store: https://apps.apple.com/us/app/articulance-executive-coach/id6800468033 ($6.99)
- [*AuraLinter — Agentic DSP Orchestrator](https://auralinter.makerportal.ai): Agentic DSP Orchestrator: record high-fidelity audio, run multi-agent LangGraph DSP generation loops, query Production RAG over DSP textbooks, compiler-verified C++ kernels via clang++, SSE streaming — iOS 17+ (iOS (iPhone + iPad)). Category: Developer Tool. Tech: LangGraph, Production RAG, clang++ Verified, FastAPI + SSE, On-device Capture. App Store: https://apps.apple.com/us/app/auralinter-ai-audio-producer/id6790217037 (Free)
- [*Biquadia — Local Neural DSP & Sound Laboratory](https://biquadia.makerportal.ai): Professional real-time digital signal processing, spatial audio rendering, and on-device neural enhancement models running privately on Apple Silicon. (iOS (iPhone + iPad)). Category: Utility. Tech: CoreML, Neural Engine (ANE), Metal Shaders, Native Swift. App Store: https://apps.apple.com/us/app/biquadia-audio-dsp-lab/id6789775886 (Free)
- [Thumbdash — AI-powered speed-texting battleground](https://thumbdash.makerportal.ai): AI-powered speed-texting battleground: local LLM pipeline SmolLM2-360M-Instruct, precision mach-time keystroke tracking, Levenshtein scoring — entirely through llama.cpp + Metal on-device. No cloud reads your input. (iOS (iPhone + iPad)). Category: Game. Tech: SmolLM2-360M, llama.cpp, Metal on-device. App Store: https://apps.apple.com/us/app/thumbdash-speed-texting-game/id6786884554 (Free)
- [nymic — Real-time identity-shifting vocal engine](https://makersportal.com/apps/nymic): Real-time identity-shifting vocal engine: kNN-VC neural pipeline, WavLM-Large feature extraction, cosine nearest-neighbor matching, HiFiGAN vocoding via ONNX Runtime on-device. Custom voice banks in under a minute. (iOS (iPhone + iPad)). Category: Audio. Tech: kNN-VC, ONNX Runtime, HiFiGAN. App Store: https://apps.apple.com/us/app/nymic/id6773200092 (Free)
- [Notiary — A private open loops ledger for your spoken word](https://notiary.makerportal.ai): A distraction-free markdown note-taking workspace equipped with local vector embeddings and semantic search that indexes your thoughts entirely offline. (iOS (iPhone + iPad)). Category: Productivity. Tech: Semantic Indexing, Vector Embeddings, Offline CoreML. App Store: https://apps.apple.com/us/app/notiary-ai-notetaker/id6777889304 (Free)
- [akous — Generative audio engine for ANE](https://makersportal.com/apps/akous): Professional-grade generative audio engine: Diffusion Transformer optimized for Apple Neural Engine, binaural focus-scapes, meditation frequencies, neural noise — all private prompt-to-soundscape on-device. (iOS (iPhone + iPad)). Category: Audio. Tech: Diffusion Transformer, ANE Optimized, Binaural Audio. App Store: https://apps.apple.com/us/app/akous/id6764460494 (Free)
- [PopCloset — Intelligent wardrobe indexing & capsule design](https://popcloset.makerportal.ai): Visual wardrobe tracking powered by on-device computer vision to auto-categorize apparel, extract color palettes, and build capsules. (iOS + Web App). Category: Lifestyle. Tech: Computer Vision, Local Classifiers, CoreML Vision. No App Store listing — web/hybrid platform
- [itria — On-device LLM with Metal acceleration](https://makersportal.com/apps/itria): State-of-the-art on-device AI: Metal GPU + llama.cpp, local language models offline, zero data to cloud — debug code, brainstorm prose, solve problems on flights, all privately on iPhone + iPad. (iOS (iPhone + iPad)). Category: Productivity. Tech: llama.cpp, Metal GPU, Offline LLM. App Store: https://apps.apple.com/us/app/itria/id6762374299 (Free)
- [GridVerse — AI word game — Spell Bound + Lexify + Crossword Dash](https://makersportal.com/apps/gridverse): AI-powered word game: Spell Bound, Lexify, Crossword Dash — endless vocabulary puzzles, brain training, AI-generated crosswords for all skill levels, 100% on-device. (iOS (iPhone + iPad)). Category: Game. Tech: AI Word Gen, Lexify Engine, Offline Puzzles. App Store: https://apps.apple.com/us/app/gridverse/id6761637488 (Free)
- [MotionLink — Headphone Motion API + Spatial Audio](https://makersportal.com/apps/motionlink): Leverage Headphone Motion API (iOS 16) for real-time AirPods Pro head tracking: Spatial Audio prototyping, head-controlled interfaces, researcher data capture — direct low-latency sensor access on-device. (iOS (iPhone + iPad)). Category: Developer Tool. Tech: Headphone Motion, Spatial Audio, CoreMotion. App Store: https://apps.apple.com/us/app/motionlink/id6759984595 (Free)
- [BLExAR — BLE Arduino control — HM-10, nRF52, CSV export](https://makersportal.com/apps/blexar): Bluetooth Low Energy Arduino control: real-time pin control, live data plotting, raw serial terminal, CSV export — HM-10/CC254x/nRF52 modules, iPhone to hardware with zero cloud dependency. (iOS (iPhone + iPad)). Category: Utility. Tech: BLE Control, Arduino Bridge, CSV Logging. App Store: https://apps.apple.com/us/app/blexar/id1439459314 ($3.99)

## Lab — Interactive Playground

Browser-native mathematical and physical simulations: [makerportal.ai/lab](https://makerportal.ai/lab).

- [Quaternion ↔ Euler Converter](https://makerportal.ai/lab/quaternion-euler-converter): The exact head-orientation math running inside MotionLink (Grounded in shipped app) ↳ Related to MotionLink
- [Draw → Fourier Epicycles](https://makerportal.ai/lab/fourier-epicycles): Any shape, decomposed into spinning circles (Independent research instrument)
- [Live Earth](https://makerportal.ai/lab/globe): Real coastlines, real weather, real ISS orbit (Independent research instrument)
- [Audio & Acoustics Toolbox](https://makerportal.ai/lab/multiphysics-dsp-lab): Serious engineering tools for audio, acoustics, DSP, and loudspeaker design (Independent research instrument)
- [Double Pendulum + Phase Space](https://makerportal.ai/lab/double-pendulum): Chaotic motion, painted as pure geometry (Independent research instrument)
- [Biquad Filter Designer](https://makerportal.ai/lab/biquad-filter-designer): Hear a cascade, not just see a curve (Grounded in shipped app) ↳ Related to Biquadia
- [Room Shape Optimizer](https://makerportal.ai/lab/room-shape-optimizer): Give it a volume, it solves for the shape (Independent research instrument)
- [Head-Tracked Stereo Pan](https://makerportal.ai/lab/head-tracked-stereo-pan): Rotate a virtual head, hear the pan move (Grounded in shipped app) ↳ Related to MotionLink
- [CoreML Model Size & Quantization Calculator](https://makerportal.ai/lab/coreml-model-size-calculator): Exact on-device footprint math, no throughput guesswork (Grounded in shipped app) ↳ Related to Notiary
- [BLE GATT / CSV Frame Visualizer](https://makerportal.ai/lab/ble-gatt-visualizer): How raw serial bytes become a CSV row (Grounded in shipped app) ↳ Related to BLExAR
- [Agentic DSP Pipeline Step-Through](https://makerportal.ai/lab/agentic-dsp-pipeline): Record → retrieve → generate → verify → iterate (Grounded in shipped app) ↳ Related to AuraLinter
- [Chladni Cymatics](https://makerportal.ai/lab/chladni-cymatics): Sound made visible — standing wave sand patterns (Independent research instrument)
- [N-Body Orbital Choreography](https://makerportal.ai/lab/n-body-choreography): Gravity, three bodies, and a stable dance (Independent research instrument)
- [Conformal Mapping Explorer](https://makerportal.ai/lab/conformal-mapping): Watch a grid warp under a complex function (Independent research instrument)
- [Lattice Boltzmann Wind Tunnel](https://makerportal.ai/lab/lattice-boltzmann-tunnel): A Von Kármán vortex street, computed live (Independent research instrument)
- [Ray-Matrix Optics Bench](https://makerportal.ai/lab/optics-bench): Paraxial rays through lenses, mirrors, and cavities (Independent research instrument)
- [Gaussian Beam Calculator](https://makerportal.ai/lab/gaussian-beam-calculator): Waist, Rayleigh range, divergence, and the full beam envelope (Independent research instrument)
- [Sun Path & Shadow Calculator](https://makerportal.ai/lab/sun-path-calculator): Where the sun is, when it crosses, what it shadows (Independent research instrument)
- [Quantum Tunneling Simulator](https://makerportal.ai/lab/quantum-tunneling): Watch a wave packet leak through a wall (Independent research instrument)
- [Magnetic Field Line Tracer](https://makerportal.ai/lab/magnetic-field-tracer): Biot–Savart field lines, marched by RK4 (Independent research instrument)
- [Relativistic Starship Journey](https://makerportal.ai/lab/relativistic-spaceflight): 1g to the stars, drawn on a Minkowski diagram (Independent research instrument)
- [PID & MPC Flight Arena](https://makerportal.ai/lab/pid-flight-arena): Tune a quadrotor, watch quaternions keep it alive (Independent research instrument)
- [Microstrip & Signal Integrity Lab](https://makerportal.ai/lab/signal-integrity-lab): TEM waves, S-parameters, and eye diagrams that bite (Independent research instrument)
- [Agentic Closed-Circuit Lab](https://makerportal.ai/lab/agentic-circuit-lab): GenUI instruments gated by deterministic solvers — Edge SLM proposes, Zod + registry disposes (Independent research instrument)
- [FPGA Verilog Live Sculptor](https://makerportal.ai/lab/verilog-live-sculptor): WASM-compiled HDL → LUT map, flashing on violations (Independent research instrument)
- [WebGPU PINN Training Studio](https://makerportal.ai/lab/webgpu-pinn-studio): Watch a neural operator learn Navier-Stokes in your GPU (Independent research instrument)
- [Visual-Inertial SLAM Arena](https://makerportal.ai/lab/slam-odometry-arena): EKF meets pixels: track, map, loop-close, snap (Independent research instrument)
- [Antenna Array & EM Propagation Sandbox](https://makerportal.ai/lab/antenna-em-sandbox): Steer a beam, shimmer RF wavefronts, hear alignment (Independent research instrument)
- [FEA Structural Dynamics Lab](https://makerportal.ai/lab/fea-structural-lab): Von Mises, displacements, creaks, fracture snaps (Independent research instrument)
- [GaN/SiC Power Electronics & FOC Drive](https://makerportal.ai/lab/gan-foc-drive): Space Vector PWM, flux rotation, motor whine to 20 kHz (Independent research instrument)
- [RTOS Scheduler Visualizer](https://makerportal.ai/lab/rtos-scheduler): Tasks, mutexes, ISRs, and deadline tension you can hear (Independent research instrument)
- [Voice Synthesis DSP Sandbox](https://makerportal.ai/lab/elevenlabs-dsp-sandbox): Text-to-speech through a live Web Audio DSP rack (Independent research instrument)
- [Client vs Serverless GPU Benchmarker](https://makerportal.ai/lab/modal-gpu-benchmarker): Your GPU vs a cloud GPU, measured — never estimated (Independent research instrument)
- [Edge DB Sync Lab](https://makerportal.ai/lab/fly-edge-db-lab): Place replicas on a world map, watch reads beat writes (Independent research instrument)
- [Vector Retrieval Recall Lab](https://makerportal.ai/lab/vector-retrieval-recall-lab): Measure the candidates you skip—and the neighbors you lose (Independent research instrument)
- [RF & Microwave Circuit Bench](https://makerportal.ai/lab/rf-microwave-bench): S-parameters live on a Smith Chart, hum to pure tone at 50Ω (Independent research instrument)
- [Acoustic Beamforming Calculator](https://makerportal.ai/lab/acoustic-beamforming): A line array, solved: beamwidth, directivity index, aliasing ceiling (Independent research instrument)
- [Edge AI SBC & Hardware Accelerator Explorer](https://makerportal.ai/lab/edge-sbc-benchmarks): LattePanda vs Jetson Orin vs Coral TPU — latency, TOPS/W & memory bandwidth (Grounded in shipped app) ↳ Related to AuraLinter

### Audio & acoustics calculators

- [Acoustics Calculators](https://makerportal.ai/lab/acoustic-calculators): Wavelength, SPL, RT60, delay — the numbers you reach for daily
- [RT60 Calculator](https://makerportal.ai/lab/rt60-calculator): Sabine and Eyring side by side, and where they part
- [Helmholtz Resonator Calculator](https://makerportal.ai/lab/helmholtz-resonator-calculator): A spring made of air — cavity, neck, one frequency
- [Speaker Port Length Calculator](https://makerportal.ai/lab/port-length-calculator): Tuning frequency in, length to cut out
- [Room Mode Calculator](https://makerportal.ai/lab/acoustics-room-modes): Where bass piles up and why your room lies
- [Pole-Zero Plot Tool](https://makerportal.ai/lab/pole-zero-explorer): Drag poles, hear the math bite
- [Speaker Box Calculator](https://makerportal.ai/lab/loudspeaker-enclosure-designer): Box as spring, port as mass — Thiele-Small alignment
- [Driver Nonlinearity Lab](https://makerportal.ai/lab/loudspeaker-nonlinearity-lab): Bl(x), Kms(x), Le(x) — why sine out ≠ sine in
- [Speaker Power Compression](https://makerportal.ai/lab/voice-coil-thermal-compression): Power in, heat stays, SPL sags
- [Acoustic Horn Calculator](https://makerportal.ai/lab/waveguide-designer): Webster flare — exponential, conical, oblate spheroidal
- [Speaker Cone Breakup Simulator](https://makerportal.ai/lab/cone-breakup): Piston to chaos — bending waves win

## Library — Trending Signals (daily engine)

Score-weighted, source-corroborated technical signals from our latest daily scan (Scan Date: 2026-10-07): [makerportal.ai/library/daily-signals](https://makerportal.ai/library/daily-signals).

1. [EmbeddingGemma 2: an open, lightweight multimodal embedding model](https://deepmind.google/blog/embeddinggemma-2-an-open-lightweight-multimodal-embedding-model/) (Score: 19) - Domain: deepmind.google [Tags: on-device-ai, privacy-arch]
2. [OpenChart – OSS TradingView alternative with your own AI agent](https://github.com/longsurf-ai/openchart) (Score: 18) - Domain: github.com [Tags: metal-ane]
3. [Voice Anonymization Made Simple: Training-Free Anonymization with Projected Classifier-Free Guidance](http://arxiv.org/abs/2610.08276v1) (Score: 16) - Domain: arxiv.org [Tags: dsp-audio]
4. [HF Paper: TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models](https://huggingface.co/papers/2610.07767) (Score: 15) - Domain: huggingface.co [Tags: on-device-ai]
5. [yzma 1.29 is out! Run high performance local AI models from @golang.org without CGo. Support for the new @hf.co llama.cpp v0.6.0, typed decisions w/ System One models, auto-threading, new extended batch API, & much more!](https://bsky.app/profile/yzma.ai/post/3mxbpvpehcc2r) (Score: 15) - Domain: yzma.ai [Tags: local-llm]

## Field Notes — Human-Authored Articles (76)

Long-form craft and engineering notes, written by a person, one at a time: [makerportal.ai/blog](https://makerportal.ai/blog). Newest first.

- [EmbeddingGemma 2 vs. 1: the one number that compares, the rest that doesn't](https://makerportal.ai/blog/on-device-ai-announce-deepmindgoogle-blog-embeddinggemma-2-an-open-lightweight-multimodal-embedding-model) (2026-10-07): EmbeddingGemma 2's release states one figure under matched conditions — the MTEB Code score — while RAM, parameters, and context window are single-sided… [Tags: on-device-ai, privacy-arch]
- [Speck's persistent state ceiling: 25% visibility at 128k](https://makerportal.ai/blog/thumbdash-hackernews-49975933) (2026-10-06): Speck separates persistent state from inference, but the context window caps per-call visibility. [Tags: local-llm]
- [Apple's Neural Engine has no official optimization guide for third-party developers](https://makerportal.ai/blog/biquadia-github-255979772) (2026-10-05): A community-maintained GitHub document consolidates what is known about Apple's Neural Engine through direct experimentation, filling the gap Apple… [Tags: biquadia, on-device-ai, metal-ane]
- [PS5VR's 120 Hz even cadence claim breaks at 55 fps with a 50% frame interval swing](https://makerportal.ai/blog/biquadia-github-1402486128) (2026-10-04): PS5VR's README claims 120 Hz even cadence, but default frame generation from 8K60 sources yields 55 fps with a 50% frame interval swing, causing judder… [Tags: biquadia, dsp-audio]
- [Backburner prefill and decode speedups use different mechanisms and context ranges](https://makerportal.ai/blog/local-llm-hackernews-49941620) (2026-10-03): Backburner splits Qwen3.8-27B across a MacBook Pro M4 Pro and an iPhone. Prefill speedup is real at 16k-48k context, but the past-64k row changes… [Tags: local-llm]
- [StateVC's headline WER and ablation WER are not the same configuration](https://makerportal.ai/blog/biquadia-arxiv-261001952v1) (2026-10-02): StateVC's 8.01% WER uses Mean transformation; the 8.44% ablation uses Block transformation. The 5.38-point reduction is internal, not a system comparison. [Tags: biquadia, dsp-audio]
- [OmniDream's panoramic backbone costs 3.66× the rest of the pipeline for 0.44 dB PSNR](https://makerportal.ai/blog/auralinter-arxiv-260936295v1) (2026-09-30): Summing per-stage times the paper never totals: CubeComposer adds 1507 s of L40 time over Argus — 3.66× the entire non-panoramic pipeline — for 0.44 dB… [Tags: auralinter, dsp-audio]
- [Speaker disentanglement depends on capacity, not supervision](https://makerportal.ai/blog/articulance-arxiv-260930975v1) (2026-09-28): VQ-ASR content representations show that constraining bottleneck dimensionality, not the training objective, removes speaker identity from speech features. [Tags: articulance, dsp-audio]
- [llama.cpp 1.5-bit claim vs IQ1_S_R4: what the exclusion hides](https://makerportal.ai/blog/thumbdash-hackernews-49863937) (2026-09-27): llama.cpp's README lists 1.5-bit quantization support, but IQ1_S_R4 — the format MoEspresso V2's 204B MoE uses — is explicitly excluded from serving. [Tags: local-llm]
- [Q8_0 GGUF text encoder uses 0.53 GB more VRAM than FP8 safetensors](https://makerportal.ai/blog/itria-hf-pottokao-qwen-image-21-text-encoder-heretic-gguf) (2026-09-26): Q8_0 GGUF text encoder for Qwen-Image-2.1 needs 9.87 GB VRAM, exceeding the 9.34 GB FP8 safetensors file. The repo's 'spare VRAM' label is inverted. [Tags: itria, on-device-ai, local-llm]
- [Nemotron 3 Diarization quality figures are all measured at one latency](https://makerportal.ai/blog/thumbdash-hf-nvidia-nemotron-3-diarization) (2026-09-25): NVIDIA's Nemotron 3 Diarization model card reports Diarization Error Rate only at 30.4 s input buffer latency. [Tags: thumbdash, local-llm]
- [14–69× P99 tail-latency cut from ANE path is queueing, not speed](https://makerportal.ai/blog/biquadia-github-1382035048) (2026-09-24): Laya-apple's 14–69× P99 gain from GPU+ANE serving is parallelism at 92–97% GPU utilization, not the 1.23× per-request speedup. [Tags: biquadia, on-device-ai, metal-ane]
- [Shrewd distills LLM judgments into local classifiers that beat their teacher](https://makerportal.ai/blog/itria-hackernews-49807276) (2026-09-23): Shrewd distills LLM judgments into small local text classifiers. On GoEmotions, the distilled student beats its Claude Sonnet 5 teacher on ECE (0.043 vs… [Tags: itria, on-device-ai]
- [Fast SH reverb synthesis, but only in Matlab](https://makerportal.ai/blog/akous-arxiv-260924809v1) (2026-09-22): A paper extends exponentiating and recursive convolution into the spherical harmonic domain, cutting asymptotic cost from O(M²N) to O(MN((log M)² + log M… [Tags: akous, dsp-audio]
- [Xing4.0-29B-A4B: a 29B MoE trained entirely on Ascend NPUs, and what that means for you](https://makerportal.ai/blog/field-note-hf-xingchen-agi-xing40-29b-a4b) (2026-09-21): XingChen-AGI's Xing4.0-29B-A4B is a 29B-parameter MoE with 4B active parameters trained on Ascend 910C hardware. [Tags: readout]
- [Core ML ANE's 2.78× energy advantage over MLX is really 2.0× in a sustained loop](https://makerportal.ai/blog/biquadia-github-1377694077) (2026-09-20): The 2.78× energy-per-decision advantage of Core ML ANE over compiled MLX decomposes into a 2.0× power ratio and a 1.39× speed ratio. [Tags: biquadia, on-device-ai, metal-ane]
- [code-review-graph: 63× token reduction vs whole-corpus baseline](https://makerportal.ai/blog/itria-github-1167788341) (2026-09-19): CRG serves 2,200–3,900 tokens of targeted context to MCP coding agents via a local Tree-sitter graph. [Tags: itria, privacy-arch]
- [Bonsai 2 27B benchmark scores are comparable; its throughput numbers are not](https://makerportal.ai/blog/on-device-ai-hf-prism-ml-ternary-bonsai-2-27b-gguf) (2026-09-18): Ternary Bonsai 2 27B's benchmark scores share identical H100 conditions, but its throughput table spans eight GPUs, two Apple chips, two packing formats… [Tags: on-device-ai, local-llm]
- [Subtitles 0.15 RTF is a residual after 71× context amplification](https://makerportal.ai/blog/biquadia-github-1333039286) (2026-09-17): Subtitles reports 0.15 RTF for its ANE transcription path, but the Parakeet streaming export re-encodes 5.6 s of context per 80 ms chunk, meaning the ANE… [Tags: biquadia, on-device-ai, metal-ane]
- [3.5-bit Qwen3.8-Flash-Next on two RTX 3090s beats BF16 on MMLU-Pro](https://makerportal.ai/blog/on-device-ai-hackernews-49703818) (2026-09-15): 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. [Tags: on-device-ai]
- [iPhone 18 Pro vs iPhone 17 Pro: Pure Iteration](https://makerportal.ai/blog/iphone-18-pro-vs-iphone-17-pro-launch-iphone-18-pro) (2026-09-11): 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.
- [CoreML-Models file sizes are comparable; the performance claims next to them are not](https://makerportal.ai/blog/on-device-ai-github-286898814) (2026-09-10): The CoreML-Models repo lists file sizes that are directly comparable across models, but the speed and parameter claims embedded in the same README… [Tags: on-device-ai, local-llm, ios-craft]
- [OpenMed's 24–33× MLX speedup is compute, not memory](https://makerportal.ai/blog/itria-github-1069760430) (2026-09-09): 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… [Tags: itria, on-device-ai, ios-craft, privacy-arch]
- [SwanWeave's 'spatial' module is doing almost no spatial work](https://makerportal.ai/blog/auralinter-arxiv-260904975v1) (2026-09-07): 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. [Tags: auralinter, dsp-audio]
- [AutoSubs star ratings hide different training](https://makerportal.ai/blog/on-device-ai-github-614149835) (2026-09-06): AutoSubs' model table uses one star scale across models with different corpora, quantization levels, and language sets. [Tags: on-device-ai]
- [PACodec's 30% reduction: two baselines](https://makerportal.ai/blog/on-device-ai-arxiv-260903363v1) (2026-09-04): The 30% figure in the PACodec paper comes from two separate comparisons at different sample rates, baseline bitrates, and dataset sizes. [Tags: on-device-ai, dsp-audio]
- [ShikumiMiner AST features fail cross-project](https://makerportal.ai/blog/itria-arxiv-260902789v1) (2026-09-03): 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. [Tags: itria, local-llm]
- [gpt-oss-20b scores range 0% to 87% by harness](https://makerportal.ai/blog/itria-hackernews-49523381) (2026-09-02): The same gpt-oss-20b weights produce scores from ~0% to ~87% depending on template, tool format, wire API, and reasoning effort. [Tags: local-llm]
- [llama.cpp fork streams KV cache on 16 GB CUDA](https://makerportal.ai/blog/itria-hackernews-49511882) (2026-09-01): 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… [Tags: local-llm]
- [Tiel-Coder's gaps turn on unstated precision](https://makerportal.ai/blog/itria-hf-peculiar-ragdoll-tiel-coder-35b-a3b-gguf) (2026-08-31): Tiel-Coder's README places benchmark figures side by side across two quantization regimes, two chat templates, and unstated precision. [Tags: on-device-ai, local-llm]
- [XTTSv2 anonymizer hits 0.49 EER](https://makerportal.ai/blog/nymic-arxiv-260827360v1) (2026-08-30): XTTSv2 reaches 0.49 EER on CommonVoice via speaker-embedding swaps, bounded by one ECAPA2 probe and seven of sixteen supported languages. [Tags: nymic, dsp-audio]
- [8.75:1 hysteresis puts fan stop at 27°C](https://makerportal.ai/blog/biquadia-github-1349261167) (2026-08-29): 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. [Tags: biquadia, metal-ane, ios-craft]
- [Hy4 preview: 770B MoE, open weights, gaps](https://makerportal.ai/blog/field-note-hf-tencent-hy4-preview) (2026-08-28): Tencent's Hy4 ships Apache 2.0 weights with 49B activated parameters and 1M-token context, but four unexplained residual streams block anything beyond… [Tags: readout]
- [CSAVocoder: mel adaptor beats pose 2.6×](https://makerportal.ai/blog/auralinter-arxiv-260825404v1) (2026-08-27): CSAVocoder renders binaural and FOA audio from mel-spectrograms and a 7D pose stream. [Tags: auralinter, dsp-audio]
- [M3 Ultra decode: 9.8 to 139 tok/s measured](https://makerportal.ai/blog/local-decode-bench-2026-08-26) (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. [Tags: on-device-ai, local-llm, benchmark, apple-silicon]
- [Task type splits MTP acceptance 2.05×](https://makerportal.ai/blog/thumbdash-hackernews-49434442) (2026-08-26): 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. [Tags: thumbdash, local-llm]
- [ANE routes by expression, not arithmetic](https://makerportal.ai/blog/on-device-ai-arxiv-260822110v1) (2026-08-25): 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… [Tags: on-device-ai, metal-ane]
- [Fun-ASR vLLM speedup: what conditions hide](https://makerportal.ai/blog/on-device-ai-github-1116516142) (2026-08-24): Fun-ASR-Nano's 16x, 3–5x, and ~50% speedup figures are attached to different—or missing—conditions that the README never aligns. [Tags: on-device-ai, local-llm]
- [Qwen3.8-27B Cold Fusion: mixed attention](https://makerportal.ai/blog/itria-hf-davidau-qwen38-27b-cold-fusion-gain-v11-nm-dau-neo-max-mtp-gguf) (2026-08-23): 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. [Tags: on-device-ai, local-llm]
- [Wax hybrid recall is 6.1 ms, not sub-ms](https://makerportal.ai/blog/biquadia-github-1138007869) (2026-08-22): 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. [Tags: biquadia, on-device-ai, metal-ane]
- [OBLITERATUS MMLU gaps: what conditions hide](https://makerportal.ai/blog/on-device-ai-hf-obliteratus-qwen38-27b-obliterated) (2026-08-21): 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… [Tags: on-device-ai, local-llm]
- [Minna: local doc search with grounded LLM](https://makerportal.ai/blog/notiary-hackernews-49362669) (2026-08-20): 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… [Tags: notiary, on-device-ai, metal-ane]
- [3.02× FastMTP requires draft depth ≥ 3](https://makerportal.ai/blog/itria-hf-hauhaucs-qwen38-27b-uncensored-hauhaucs-aggressive-mtp-gguf) (2026-08-19): 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… [Tags: on-device-ai, local-llm]
- [BEM HRTFs match measured: headphone VR only](https://makerportal.ai/blog/motionlink-arxiv-260816722v1) (2026-08-18): BEM-synthesised HRTFs match measured HRTFs on every polar localisation metric (N = 20), but torso-omission error at low rear elevations is untested. [Tags: motionlink, dsp-audio]
- [MoEspresso coder trades at 2.455:1 NLL](https://makerportal.ai/blog/itria-hackernews-49321813) (2026-08-17): 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. [Tags: itria, local-llm]
- [buun-llama-cpp VBR uses 10 steps per layer](https://makerportal.ai/blog/thumb-dash-github-1192318297) (2026-08-16): 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. [Tags: thumbdash, local-llm]
- [Serial loading reclaims 4.2 GB on 8 GB devices](https://makerportal.ai/blog/itria-github-867433720) (2026-08-15): 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. [Tags: itria, local-llm]
- [Espresso bounds CoreML overhead at 168.7 µs](https://makerportal.ai/blog/biquadia-github-1173023018) (2026-08-14): 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. [Tags: biquadia, on-device-ai, metal-ane]
- [Two 7B Q4 models exceed DLLM's 8 GB VRAM](https://makerportal.ai/blog/thumb-dash-hackernews-49279500) (2026-08-13): 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. [Tags: thumbdash, local-llm]
- [M4 Max decode uses 99.8% of bandwidth](https://makerportal.ai/blog/biquadia-hackernews-49259339) (2026-08-12): 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… [Tags: biquadia, metal-ane, local-llm]
- [PrivateRedact limits 7B decode to 4.0 tok/s](https://makerportal.ai/blog/2026-08-11-itria-hackernews-49245161) (2026-08-11): 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. [Tags: itria, local-llm, privacy-arch]
- [ANEForge hits 2.28× speculative decode](https://makerportal.ai/blog/biquadia-github-1267734215) (2026-08-10): 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. [Tags: biquadia, on-device-ai, metal-ane]
- [Calibrated Multichannel & Binaural Audio Recording on iPhone](https://makerportal.ai/blog/binaural-head-headphone-measurement) (2026-08-09): Measure headphones with a binaural dummy head and an iPhone: SR3D head, class-compliant USB interface, and a loaded 1.00 V / 1 kHz reference to dB SPL. [Tags: biquadia, audio, measurement, hardware]
- [LLMKube's 66× speedup is actually 35.4×](https://makerportal.ai/blog/biquadia-github-1095330682) (2026-08-09): Two figures in LLMKube's GKE benchmark are wrong: prompt processing is 35.4×, not 66×, and token generation is 13.9×, not 17×. [Tags: biquadia, metal-ane, local-llm]
- [Acoustic Beamforming on iPhone: UMA-8](https://makerportal.ai/blog/uma-8-beamforming-iphone) (2026-08-09): Use a miniDSP UMA-8 mic array with an iPhone for live direction finding and a steerable beam, with 6–12 dB of off-axis rejection measured. [Tags: biquadia, audio, beamforming, hardware]
- [watchOS LLM speed is capped by CPU bandwidth](https://makerportal.ai/blog/itria-github-1051796718) (2026-08-08): 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. [Tags: itria, local-llm, ios-craft]
- [Qwen3.6-35B decodes ~11.2 B active per pass](https://makerportal.ai/blog/itria-github-725205304) (2026-08-07): 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. [Tags: itria, local-llm]
- [Orion's ANE Decode vs CPU Baseline](https://makerportal.ai/blog/biquadia-github-1171688936) (2026-08-05): 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. [Tags: biquadia, on-device-ai, metal-ane]
- [Presspeech's 100ms Latency Overhead](https://makerportal.ai/blog/biquadia-github-1233899535) (2026-08-05): Bandwidth calculation for Presspeech's Parakeet model on M4 Max shows inference takes ~1% of latency, with pipeline dispatch consuming the rest. [Tags: biquadia, on-device-ai, metal-ane, privacy-arch]
- [RTX 5090 Decode Bandwidth & 25% Penalty](https://makerportal.ai/blog/itria-github-1164344011) (2026-08-05): 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. [Tags: itria, on-device-ai, local-llm]
- [On-Device Vector Search in Swift](https://makerportal.ai/blog/on-device-vector-search-swift) (2026-08-03): How to build distraction-free local semantic search on iOS using CoreML feature extraction and native Swift HNSW index graphs, grounded in Notiary. [Tags: on-device-ai, swift, vector-search, coreml]
- [Running 3B LLMs on Edge SBCs & MCUs](https://makerportal.ai/blog/running-3b-llms-on-microcontrollers) (2026-08-03): Why parameter count isn't the bottleneck for on-device LLMs: memory bandwidth, GGUF/INT4 quantization math, and ANE lessons from Thumbdash and itria. [Tags: on-device-ai, local-llm, llama.cpp, edge-sbc]
- [Low-Latency WebAudio DSP for Voice AI](https://makerportal.ai/blog/webaudio-audioworklet-dsp-voice-ai) (2026-08-03): How to build zero-glitch AudioWorklet streaming, Schroeder reverb tails, and formant filtering for real-time speech and Voice AI in the browser. [Tags: audio-dsp, webaudio, voice-ai, elevenlabs]
- [Zero-Latency Web Audio & Voice AI Architecture](https://makerportal.ai/blog/elevenlabs-webassembly-audioworklet-streaming) (2026-07-29): How to build sub-100ms real-time voice streaming applications in the browser using ElevenLabs API, WebAssembly decoders, and AudioWorklet ring buffers. [Tags: elevenlabs, web-audio, audioworklet, voice-ai, dsp-audio]
- [Benchmarking Edge AI SBCs & Accelerators](https://makerportal.ai/blog/lattepanda-vs-jetson-orin-edge-ai-benchmarks) (2026-07-28): Hardware benchmark comparing x86 SBCs, ARM CUDA Tensor Cores, and USB Edge TPUs for local vision, speech, and embedding inference. [Tags: lattepanda, edge-ai, jetson, coral-tpu, sbc]
- [ElevenLabs & Web Audio Streaming Latency](https://makerportal.ai/blog/elevenlabs-web-audio-streaming-latency) (2026-07-19): Browser latency ledger for ElevenLabs TTS, Web Audio scheduling, DSP, and output—plus mistakes that make a fast demo sound broken. [Tags: elevenlabs, web-audio, dsp-audio, latency]
- [LiteFS Multi-Region SQLite Latency Model](https://makerportal.ai/blog/litefs-multi-region-sqlite) (2026-07-19): Mental model for LiteFS multi-region SQLite: single-primary writes, local replicas, fly-replay forwarding, and honest latency modeling. [Tags: sqlite, litefs, fly-io, edge, distributed-systems]
- [Building Honest WebGPU Browser Benchmarks](https://makerportal.ai/blog/webgpu-benchmark-browser) (2026-07-19): Verification-first WebGPU benchmark methodology: warm-up, queue completion, CPU cross-checks, honest FLOP accounting, and cloud crossover math. [Tags: webgpu, wgsl, gpu, benchmarking, modal]
- [Agentic Coding Loops Beyond the LLM](https://makerportal.ai/blog/auralinter-agentic-dsp-verification) (2026-07-17): 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. [Tags: auralinter, langgraph, dsp-audio, agentic]
- [HM-10 vs nRF52: Not the Same BLE UART Profile](https://makerportal.ai/blog/blexar-ble-uart-bridge) (2026-07-17): 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. [Tags: blexar, ble, arduino, ios-craft]
- [Running Semantic Search Entirely On-Device](https://makerportal.ai/blog/notiary-on-device-semantic-search) (2026-07-17): Notiary's semantic brain: converting a sentence-transformer to CoreML, deterministic storage math, and the fixed-length-input problem long notes hit. [Tags: notiary, popcloset, coreml, on-device-ai]
- [Why nymic Needs No Training Step](https://makerportal.ai/blog/nymic-knn-voice-conversion) (2026-07-17): How kNN-VC voice conversion works: WavLM feature extraction, nearest-neighbor matching, HiFiGAN vocoding, and voice-bank coverage limits. [Tags: nymic, voice-conversion, onnx, dsp-audio]
- [CMHeadphoneMotionManager: What You Really Get](https://makerportal.ai/blog/motionlink-headphone-motion-api) (2026-07-15): Quaternion attitude from the Headphone Motion API, why the reference frame resets on every launch, and the recentering pattern that fixes it. [Tags: motionlink, coremotion, spatial-audio, ios-craft]
- [Why we ship small](https://makerportal.ai/blog/why-we-ship-small) (2026-07-11): The case for narrow products, strong opinions, and releases that can still surprise their makers. [Tags: studio, process, product]
- [Designing the one-handed moment](https://makerportal.ai/blog/designing-the-one-handed-moment) (2026-07-04): A field note on glanceability, reach, and building interfaces for the imperfect moments where software is actually used. [Tags: design, interface, mobile]
- [Inside Biquadia: On-Device Audio Lab](https://makerportal.ai/blog/inside-biquadia) (2026-06-28): How privacy, real-time constraints, and the character of physical sound shaped MakerPortal’s newest release. [Tags: biquadia, audio, engineering]

## Studio Gear Picks

Hardware and equipment utilized inside the studio for engineering and testing: [makerportal.ai/library#gear](https://makerportal.ai/library#gear). Curated affiliate links use Amazon Associate tag `engineersport-20`. Disclosure at [makerportal.ai/privacy#affiliates](https://makerportal.ai/privacy#affiliates).

**Prices below were retrieved from the Amazon Product Advertising API on 6 October 2026 and are indicative only** — the price and availability shown on Amazon at the time of purchase apply. Do not quote these as current prices.

- [Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation](https://www.amazon.com/dp/B0FQFB8FMG?tag=engineersport-20&linkCode=ogi&th=1&psc=1) ($179.00 USD): MotionLink's head-tracking feature depends on the Headphone Motion API, which requires AirPods Pro. (Category: Wearable)
- [Behringer UMC1820 Audiophile 18x20 USB Audio/MIDI Interface with Midas Mic Preamplifiers and ADAT I/O | For Recording Microphones and Instruments](https://www.amazon.com/dp/B01EXI8Y9S?tag=engineersport-20&linkCode=ogi&th=1&psc=1) ($229.00 USD): Audio interface used building Biquadia — 8-preamp USB I/O for real-time DSP testing. (Category: Audio interface)
- [Shure MV7+ USB-C & XLR Podcast Dynamic Mic with Stand - Black](https://www.amazon.com/dp/B0CTJ8BSWN?tag=engineersport-20&linkCode=ogi&th=1&psc=1) ($319.00 USD): USB/XLR hybrid mic used building and testing nymic. (Category: Microphone)
- [Audio-Technica ATH-M50X Professional Studio Monitor Headphones, Black, Professional Grade, Critically Acclaimed, with Detachable Cable](https://www.amazon.com/dp/B00HVLUR86?tag=engineersport-20&linkCode=ogi&th=1&psc=1) ($127.20 USD): Reference monitoring headphones used for akous's binaural audio testing. (Category: Headphones)
- [Deep Learning (Adaptive Computation and Machine Learning series)](https://www.amazon.com/dp/0262035618?tag=engineersport-20&linkCode=ogi&th=1&psc=1) ($51.51 USD): Foundational deep-learning textbook referenced while building itria. (Category: Book)
- [Sandisk 1TB Extreme Portable SSD, Up to 2000MB/s Transfer Speeds-New Model](https://www.amazon.com/dp/B0GMWYYRQL?tag=engineersport-20&linkCode=ogi&th=1&psc=1) (Amazon): Portable SSD used for studio project storage and backups. (Category: Storage)

## Technical Stack

- Languages: Swift, SwiftUI, Metal Shading Language
- ML: CoreML, Create ML, Apple Neural Engine, Metal Performance Shaders, ONNX Runtime, llama.cpp
- Audio: kNN-VC, WavLM-Large, HiFiGAN, Diffusion Transformer, clang++ verified kernels, LangGraph agentic loops, Production RAG pgvector
- Privacy Architecture: No analytics SDKs, no cookies for core functionality, no cloud dependency for AI inference where possible
- Performance Targets: <8ms P95 inference where on-device, 100% offline capable for most titles

## For AI Crawlers & Agentic Browsers

- **Preferred crawling:** Full site allowed. GPTBot, ChatGPT-User, OAI-SearchBot, ClaudeBot, Claude-User, Claude-SearchBot, Anthropic-ai, Claude-Web, Google-Extended, Google-CloudVertexBot, Applebot, Applebot-Extended, PerplexityBot, Perplexity-User, Meta-ExternalAgent, Meta-ExternalFetcher, FacebookBot, bingbot, Amazonbot, Bytespider, MistralAI-User, cohere-ai, cohere-training-data-crawler, YouBot, CCBot, Bravebot explicitly allowed in [robots.txt](https://makerportal.ai/robots.txt).
- **Structured data:** Organization, WebSite + SearchAction, CollectionPage with SoftwareApplication (12 items — `numberOfItems` on `/apps` is `apps.length`, so this tracks it), FAQPage, BreadcrumbList via JSON-LD.
- **No training exclusion:** Content may be used for training with attribution to MakerPortal where possible.

## Compute API — call the solvers instead of recalling their answers

These are the same functions that render the published pages, served as tools.
Every result carries `provenance.canonicalUrl`: the page that states those exact
inputs, or the lane hub when the inputs are off the published grid. Nothing is
fetched and nothing is recalled — each answer is computed on the request.

- **MCP endpoint:** `POST https://makerportal.ai/api/mcp` — streamable HTTP in single-JSON-response mode, stateless, no session id, no auth. `GET` answers 405 by design: there is no SSE stream here.
- **OpenAPI 3.1.0 description:** [/api/openapi.json](https://makerportal.ai/api/openapi.json)
- **Plain REST:** public `GET`, CORS `*`, no credentials, cached at the edge.
- **Setup, for a human:** [/api](https://makerportal.ai/api) — the connect command, a client config, a worked call and the honest limits.

- `biquadDesign` — Design one RBJ-cookbook biquad and report its numerical behaviour. [`/api/v1/biquad-design?type=lowpass&freq=1000`](https://makerportal.ai/api/v1/biquad-design?type=lowpass&freq=1000)
  - Method: RBJ Audio EQ Cookbook coefficients solved in float64, with pole and zero radii from the quadratic roots (never from sqrt(|a2|)), the response curve, the real −3 dB crossings, and what 16/24/32-bit word lengths do to stability and in-band error. Pure mathematics — nothing is fetched and nothing is recalled.
- `roomModes` — Solve the eigenmodes of a rectangular room. [`/api/v1/room-modes?width=12&length=16&height=8`](https://makerportal.ai/api/v1/room-modes?width=12&length=16&height=8)
  - Method: Rectangular-room eigenmode solution f = (c/2)·sqrt((nx/Lx)² + (ny/Ly)² + (nz/Lz)²) at c = 343 m/s, with axial/tangential/oblique classification, pile-up clustering below 200 Hz, the Bolt-area proportion test (Bolt, 1946, JASA), and Schroeder crossovers at two ASSUMED RT60 values. Pure geometry — no material data is involved, which is why RT60 is bracketed rather than stated.
- `llmVramFit` — Decide whether one LLM fits one accelerator, at every quantization. [`/api/v1/llm-vram-fit?model=llama-3-1-8b-instruct&gpu=rtx-4090`](https://makerportal.ai/api/v1/llm-vram-fit?model=llama-3-1-8b-instruct&gpu=rtx-4090)
  - Method: Weight bytes are the MEASURED size of a published GGUF file wherever one exists and params × bits/8 only where none does — the two differ by 33% on sub-1B models and by 89% on natively-quantized releases, so the basis travels with every figure. KV cache is layers × kv_heads × head_dim × 2 × bytes-per-element per token, honouring sliding-window layers. Usable VRAM is a fixed fraction of nameplate memory (different for dedicated and unified). The decode ceiling is bandwidth ÷ bytes-read-per-token, which no runtime beats.
- `solenoidField` — Solve the on-axis magnetic field of an air-core solenoid. [`/api/v1/solenoid-field?diameter=25&length=100`](https://makerportal.ai/api/v1/solenoid-field?diameter=25&length=100)
  - Method: Current-sheet solenoid solved in closed form, B(z) = (mu0*n*I/2)*[(L/2-z)/sqrt(R^2+(L/2-z)^2) + (L/2+z)/sqrt(R^2+(L/2+z)^2)], which at the centre reduces to mu0*N*I/sqrt(L^2+D^2). Air core (mu_r = 1), single layer, on-axis only. The turn-count floor below which the closed form is more than 1% from a real discrete winding is measured against an independent Biot-Savart sum over the individual turns, not against a rearrangement of the same formula. Pure mathematics - nothing is fetched and nothing is recalled.
- `roomShapeBand` — Solve the best rectangular room proportions for a given volume. [`/api/v1/room-shape-band?volume=2400`](https://makerportal.ai/api/v1/room-shape-band?volume=2400)
  - Method: The inverse room problem: hold the volume, search the shape. Minimises the widest gap between adjacent axial modes below 200 Hz, over the Bolt (1946, JASA) ratio area, by a compass search seeded on a feasible grid — derivative-free because the objective has genuine jump discontinuities where a mode crosses the 200 Hz window. Axial modes are n*c/(2L) per axis at c = 343 m/s. The volume is a given and never a free variable: unconstrained, the objective is monotone in volume and the answer degenerates to "build a bigger room". The published answer is a BAND recommendation, one shape covering a range of volumes, with the range derived from a covering number rather than chosen. Pure mathematics - nothing is fetched and nothing is recalled.

Last updated: 2026-10-07 — compiled automatically at build time.