←
Manolis Nikiforakis1,384 posts
Manolis Nikiforakis avatar

Manolis Nikiforakis

@nikil511

IoT + Web3 craftsman, CEO @WeatherXM

◷ Joined November 2010▧ 121 Photos and videos
Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Nov 2

Most AI agents can tell you the forecast. This one looks at the sky outside my window and measures the actual weather on my roof. We ported @Meta’s open-source Muse Gadget SDK to the @WeatherXM D1 (WG1200), turning a weather station into a Physical AI test bench: 📡 Live rooftop Sub-GHz LoRa sensors (sensors.read) 👁️ Eyes on the sky via local LAN IP cameras (lan.fetch) 🎨 A 480×480 generative canvas (display.draw_file) 🔄 Reversible Dual-Boot (stock WeatherXM OS stays untouched) Open-source repo & 1-click Web Flasher 🧵👇

Post attachmentPost attachment
Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Nov 2

hey @anantn quick heads-up on two things I've put together for Muse gadget ESP32: Web-based Flasher & BLE Provisioner (muse.weatherxm.com): To make it frictionless for the thousands of WeatherXM station owners to try Muse alongside their stock weather station firmware. I've built a browser-based USB flasher + Web Bluetooth provisioning flow (with a small Worker proxy for token minting/VM lookup). The BLE/token setup is admittedly a bit hacky right now — hope you don't mind us shipping it for our community testing! LAN.fetch on ESP32 (PR #299): We also opened PR #299 implementing lan.fetch on the ESP32 so the gateway can pull snapshots from local LAN sky cameras (and local HTTP endpoints) and forward the images to Muse for live visual sky/weather analysis. I hope this was within your plans for ESP32. Would love your thoughts when you get a chance to take a look! The AI + hyper local weather usecases are endless and amazing!

Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Nov 1

We ported Muse to the @WeatherXM D1 rooftop LoRa weather station, and upstreamed lan.fetch in PR #299 so any ESP32 Muse gadget can pull local LAN sky-camera snapshots. Single-station Physical AI works today, with cryptographic proofs on data provenance in our case. The real unlock is an agentic "SETI@home" for weather nowcasting: neighboring Muse agents discovering each other, tasking synchronized sky captures, and grading forecasts against rooftop weather readings. @alexandr_wang when will the Muse agent-to-agent substrate be ready so we can change the face of meteorology forever?

Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Nov 1

6/ Want to experiment with Physical AI on real outdoor instruments now!? We ship globally, like our network ! Use code MUSE50 for $50 off per station on the WeatherXM D1 Wi-Fi ($199 → $149 Network Edition)! 🛒 weatherxm.com/product/wb1200… Get two or more to test multi-agent coordination today! And subscribe for off-grid LoRa Mesh multi-hop that is comming next: crowdsupply.com/weatherxm/we…

Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Nov 1

5/ Where this goes next: Self-Assembling Physical AI Nowcasting 🌍 1️⃣ Discovery & Mesh Rendezvous: Stations within a region, each run by its own independent Muse agent, find each other via a signed rendezvous and advertise what they can observe: camera bearing, live sensors, scored history. No central planner. 2️⃣ Peer-to-Peer Tasking: "Capture the sky at T+90s, signed—I'll do the same for you." Two synchronized sky views + wind vectors → stereo cloud height, optical flow, and a shadow-arrival countdown drawn on the neighboring rooftop's D1 screen. Agentic nowcast! 3️⃣ Graded by Nature: Every prediction is scored against physical ground truth—did the rooftop solar irradiance (W/m²) actually drop at the predicted minute? "Insufficient geometry" is a valid answer, and error bars get published next to the wins. Agents get better in every turn. 📡An agentic SETI@home for weather nowcasting! Instead of just donating idle CPU cycles, Muse AI agents contribute live rooftop telemetry, night-star auto-calibrated sky-camera views, and peer-to-peer coordination.

Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Nov 1

4/ Already own a WeatherXM D1 Wi-Fi station? Create a Muse account and connect your station. And you don't have to sacrifice your station to try this. Our dual-boot signed layout installs Muse into an isolated muse_nvs partition. Your stock WeatherXM firmware, station identity, X.509 hardware certificates, and NVS remain 100% untouched and you can revert. Flash directly from Chrome/Edge, no toolchain required: ⚡ weatherxm.github.io/muse-gad…

Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Nov 1

3/ The 480×480 TFT screen of D1 is a canvas for Muse, not a static dashboard. Ask Muse (or swipe down on the WG1200 touchscreen), and it turns your station's exact live rooftop readings into a bespoke 480×480 weather card on the fly: • Cyberpunk Neon • Isometric Acropolis rooftop diorama • Ukiyo-e woodblocks & Studio Ghibli • Dismissible storm alerts with audible buzzer chimes (wg1200.notify) Same physical instruments. New generative UI every time adjusted to your taste.

Post attachmentPost attachmentPost attachment
Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Nov 1

2/ Why pair a sky camera with a weather station and a LoRa radio? A camera alone sees clouds, but doesn't know wind gusts, pressure drops, or cloud-base altitude. Sensors alone can't see a storm cell building on the horizon. On the WG1200, the Semtech SX1262 LoRa radio decodes live outdoor weather packets every 20s—temperature, humidity, wind speed & gusts, rain rate, UV, and solar irradiance. And with multi-hop D2 Meshtastic / MeshCore LoRa mesh field sensors and agents can relay weather ground truth across 100s of km off-grid ! crowdsupply.com/weatherxm/we…)

Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Nov 1

1/ Cloud AI agents are normally blind to your local network unless you open ports, pay for a camera/weather cloud, or run a Raspberry Pi bridge. So we built lan.fetch + device.discover directly into the ESP32-S3 firmware on the WeatherXM WG1200. Here, Muse reaches through the WG1200 to grab a 2560×1440 Athens sky snapshot from a local $30 Sonoff camera running open-source Thingino, analyzes the clouds in its VM alongside live rooftop station telemetry, and draws the frame straight back onto the 480×480 screen as confirmation.

Post attachmentPost attachment
Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Nov 1

Most AI agents can tell you the forecast. This one looks at the sky outside my window and measures the actual weather on my roof. We ported @Meta’s open-source Muse Gadget SDK to the @WeatherXM D1 (WG1200), turning a weather station into a Physical AI test bench: 📡 Live rooftop Sub-GHz LoRa sensors (sensors.read) 👁️ Eyes on the sky via local LAN IP cameras (lan.fetch) 🎨 A 480×480 generative canvas (display.draw_file) 🔄 Reversible Dual-Boot (stock WeatherXM OS stays untouched) Open-source repo & 1-click Web Flasher 🧵👇

Post attachmentPost attachment
Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Oct 8

Not everyone joined the network for the crypto rewards. Many people utilize the weather forecast they get with their station. As for the NFTs, you have no idea how hard the team and my self worked to make targeted rollouts economically sustainable. blog.weatherxm.com/targeted-…

Stretches with 0 communication. A token approaching 0 value. An outdated and non-functioning explorer. A crowd-funded station deployment via NFTs that never went live. Stations offline as there's no operational or maintenance incentive. How do you think this ends?

Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Oct 8

“WEATHER PORN.” 🌦️ That’s how one meteorologist at the World Meteorological Technology Expo described FACT. And yesterday there was a particularly satisfying plot twist: Michael Bührer from @meteoblue presented @WeatherXM FACT on stage. 😄 We’re evaluating 40+ forecasting models against observations from 8,000+ WeatherXM ground stations, across the entire 2025 dataset. 1.83 BILLION hourly forecasts! Global rankings. Country rankings. Per-station results. Temperature, humidity, wind and rain. Every forecast horizon — same day, +1, +2, +3, +4, +5 and +6 days ahead. One important caveat: WeatherXM MOSAIC and meteoblue mLM sit at the top, but they’re not playing exactly the same game as the other models. They use station observations as part of their post-processing/training, while the other models are evaluated without knowledge of our stations. So this isn’t meant to prove that MOS “beats” global weather models. It shows something more commercially useful: 👉 What accuracy should a customer actually expect at a physical location when buying an end-to-end observation + forecasting solution from us? The rest of the models compete on equal terms. And the results are spicy. 🌶️ fact.weatherxm.com #MTX26 #TECO2026

Post attachment
Manolis Nikiforakis avatar
Manolis Nikiforakis@nikil511·Oct 6

If anyone wants to talk forecast accuracy tracking and which is the best model per variable, horizon for every single location we have a weather station, I am around. We published 2025 results comparing 40+ models, check it out at: fact.weatherxm.com @WeatherXM

Post attachmentPost attachment

Only 10 minutes until the doors open for the exhibition floor of MTX 2026 👏 @WMO Get ready to head over❗#MTX26 #TECO2026 #MetTechWorldExpo

Gerasimos Tsiamalos avatar
⟳ Manolis Nikiforakis retweeted
Gerasimos Tsiamalos@tsiger·Oct 3

Hey @Starbucks if you ever decide to attack the Greek market and take over with a new name, call it Aeroclano. Thank me later.

The year was AD 2026. Starbucks discovered the Freddo Espresso and named it some silly shit, Aerocano. There’s one thing they’ll never have that the Greeks have mastered. Finesse. Say no to Starbucks and support your local businesses everyone.

Post attachment