Dozens of repos later I am somehow doing neuroscience, and every one of them still traces back to the coop.

🔗 The neuroscience: a Parkinson's vulnerability predictor

👁️ My own model, Vision Narrator 0.8B · A LoRA fine-tune of Qwen3.5-0.8B on VizWiz photos taken by blind people (CC BY 4.0), taught by a 27B teacher. It was the app's first describer.

🔗 App Store · Google Play · Source · ▶ Watch it see the world, fully offline (90 sec) · 🎙️ Episode 307 · Vision Narrator 0.8B on Hugging Face

📋 The test · 27 photos of what blind users point a phone at: medicine labels, money, curbs, stairs, signs, a dark street. Every answer was shuffled and judged with the model names hidden, and every safety miss cost 2 points. Best possible score: 96.

🏆 The result · Liquid AI's LFM3-Omni-2B beat my own Vision Narrator 0.8B on 20 of 27 photos, with 1 safety miss to my model's 5. So Liquid's model is what RealTime AI Cam ships now.

🔬 The surprise · I fine-tuned Liquid's model on my blind-user data and it got worse, 17.5 points worse at 1.1 GB. It ships exactly as Liquid released it.

📦 Shrinking it for phones · I quantized it myself, from 2 GB down. From 870 MB down the scores drop fast, and by 640 MB both 2B models fell apart, so 735 MB is the smallest that held up.

🔍 The detector rematch · In a fair rematch on 150 photos, YOLOE-26 with 4,585 words lost to the app's 601-object detector: lower recall (0.15 vs 0.23), about 3x slower, and 48% of its confident labels outside the 601 were made up. The 601 detector kept its job.

📱 In the apps now · iPhone 1.3.6 runs LFM3-Omni-2B at 735 MB, fully offline. Android 1.24 already picks from the four levels on its own.

⏭️ Next update · Version 1.4.0 brings the four levels to iPhone. A quick speed check on first launch picks the most accurate level the phone can run in time.

📊 The full benchmark, every round and every number

🧩 lfm-tool-calls · A parser and OpenAI compatible proxy that turns Liquid AI LFM tool calls into standard ones.

🌐 Browser Agent · Local MLX models driving Chrome via DevTools Protocol. Handles cross-origin iframes, Shadow DOM and ProseMirror.

🗣️ NarrateClaude + Ohm · Talk to your Mac, hear it reply in your own cloned voice. 100% private, works on a plane.

👁️ nemotron-omni-mlx · NVIDIA's Nemotron Omni had no open Apple Silicon runtime for its vision and audio towers, so I wrote one: 23/23 parity tests against NVIDIA's PyTorch reference, 67.7 tok/s with an image and 110.5 tok/s with audio on an M5 Max.

🍎 Upstream in Apple's MLX · Muse Glimmer (Meta) text model support merged into ml-explore/mlx-lm (+302 lines). Thanks to @zcbenz for the review and merge.

🖥️ Trinidad Head · My own terminal for Windows and Mac: glass windows with a neon glow, a different color for every window, your own questions easy to spot in an AI chat.

Building lately (the week of Nov 2, numbers as of Nov 2, 2026)

🔨 claude-code-local · Claude Code Local now runs on Linux and Windows through a portable llama.cpp backend, with installers for both.

🔨 browser-agent · The browser agent broker can now record browser jobs as scrubbed training trajectories if you opt in.

🔨 agent12 · Agent12 now takes community results, with a submission validator, a model picker and Will it fit my Mac pages.

🔨 lfm-tool-calls · Published a parser and OpenAI compatible proxy that turns Liquid AI LFM tool calls into standard ones.

🔨 body-movin-games · Built Body Movin, full body motion games for any TV where your phone is the controller.

🔗 claude-code-local: 3,400+ stars, 648 forks · Awesome Claude Code · Browser Broker · Browser Agent (▶ 56-second demo) · NarrateClaude + Ohm · nemotron-omni-mlx · MLX pull request · @zcbenz · Agent-12 · @galashko · lfm-tool-calls · Trinidad Head

🐝 Hive Strike · One bee against sixteen worlds, sixteen bosses and seventy-five kinds of bug. Every sprite, background, animated scene and music bed generated locally on one Mac, no cloud, no stock art.

📱 Every app, both stores. All 5 are live on the App Store and Google Play. Private by default, pay once (or nothing), no subscriptions. Android source ships in each repo's android/ folder.

📱 5 apps on the App Store and Google Play, 2,821 downloads so far (1,948 App Store, 873 Google Play), 56 new on iPhone in the last week.

🍎 App Store: Song Forge · RealTime AI Cam · RealTime Space · Brainforest · Hive Strike
🤖 Google Play: Song Forge · RealTime AI Cam · RealTime Space · Brainforest · Hive Strike

Updated automatically Nov 6, 2026 from Apple's sales reports and Google Play's install reports. Downloads are first-time App Store downloads (no updates or re-downloads) plus Google Play installs; Google runs about two weeks behind.

New iPhone downloads per week, last 8 weeks: 90 → 108 → 92 → 140 → 75 → 40 → 51 → 56 (+5 vs the week before).

🌱 Here's where all of this is headed. You hand a robot a bag of seeds, show it where the hose is, and a few months later there are vegetables on the table.

RealTime AI Cam is the eyes

Claude Code Local is the brain · The brain without a cloud.

The Cemani robot is the first body · A 1 kW autonomous tank robot you text from your phone: vision on a Jetson, dual PTZ cameras, 3D LIDAR, and it pulls a cart of firewood.

🧩 One system, three machines · M5 Max MacBook Pro · 128 GB for heavy compute, an M4 Pro Mac mini · 64 GB that never sleeps (131 background jobs: watchers, servers, tunnels), and a Contabo VPS as the public gateway.

🔒 AirGap · The same stack runs with the wifi off, for firms whose data cannot touch the cloud.

🔗 Cemani Homestead Robot · ▶ 53-second demo · VisionBuilder · AirGap · info@nicedreamzwholesale.com

🛒 What pays the bills · Divine Tribe, vaporizer hardware since 2013, plus three more stores.

📬 matt@ineedhemp.com · 📚 Case studies · HQ · Divine Tribe · three more stores · LinkedIn