Qalarc / Guides

Measured how-tos, not marketing.

By Qalarc (Sydney) · every guide documents a system we actually run · llms.txt
Everything here was built and measured in-house first — the numbers, hardware tables and honest gaps are from our own studio stacks, not benchmark theater.

Build with AI

How to build an AI character or companion with face and voiceLLM, persona, memory, voice, avatar, lip-sync — plus the interruptibility lesson from five generations of characters.Read the guide → Face AI in the browser — recognition, demographics, heart rate, no uploadMediaPipe + ONNX on WASM: the stack that works, the privacy model, and rPPG heart rate from a webcam.Read the guide → How to use AI to listen to what is said on radiosSDR + Whisper + local AI: a working open pipeline, hardware table and legal notes.Read the guide → Free small business website — the honest options (Australia, 2026)What free actually covers, which parts never are, and the built-for-you path for Australian tradies.Read the guide →

Run AI locally

Local Model Atlas — every LLM, image and audio model we actually runMeasured, installed-and-used catalogue: GLM, Qwen, Ornith, FLUX, Wan 2.2, ACE-Step, F5-TTS — with honest use-notes.Read the guide → GLM 4.7 vs Qwen3.8 vs Ornith 1.5 — measured on one mini PCSame-box, same-night head-to-heads: tokens/s, context depth and concurrency costs on 128GB unified memory.Read the guide → Strix Halo for local AI — measured reality (Ryzen AI Max+ 395 / 8060S)48-52 tok/s on MoE models, the honest 70B correction, vs DGX Spark and Mac Studio, config notes.Read the guide → Offline local AI media generation — image, video, music on your own hardwareHonest hardware tiers, the model-download reality, and a working ComfyUI-based stack.Read the guide → Monitor your Selectronic SP-PRO locally — no cloud, no subscriptionThe local HTTP endpoint, polling architecture, and the free SolarScope Android app.Read the guide →

Ship & measure

How to cut AI coding agent token costsWhere spend actually goes (tool output, not chat) and the concrete levers — with open-source QTK.Read the guide → What is QTK (Qalarc Token Killer)?Deterministic token compression for AI agents: 10.75M tokens eliminated over 17 days, no LLM, sub-millisecond.Read the guide → Analytics without cookies — first-party, aggregate, no consent bannerViews, opens, engagement and uniques with zero cookies and zero cross-site tracking: the architecture this site runs.Read the guide →