Lonboard Adds Native Cloud-Optimized GeoTIFF Support

Visualize massive amount of cloud optimized geotiffs in your notebook. Native Cloud-Optimized GeoTIFF support in Lonboard! https://lnkd.in/gt8ZVTC9 - Stream massive COGs on demand - Full control over image rendering: apply colormaps, change band combinations, use ML models - Read COGs from anywhere, including local files and private data (via Obstore)

New interactive way to insert a different plot type into an existing graph https://lnkd.in/egR6S4eD and plot against right Y axis to get a Double Y graph. A common example is adding line plot to a stacked column plot. And set it to plot against right Y axis to create a double Y graph. Learn more and download a free trial at https://lnkd.in/edn3zu6V #originpro2026 #originpro #originlab #DataAnalysis #DataVisualization #scientificgraphing #doubley #stackedcolumns #stackedbars

Vector search is not enough for RAG. It finds chunks that are semantically close to your query, but close ≠ relevant. You often end up retrieving chunks that mention the topic… without actually answering the question. That’s where reranking changes everything. Instead of embedding query and chunks separately, a cross-encoder scores them together: “Does this chunk actually answer this query?” A much stronger relevance signal. What I implemented: → Vector search over all chunks → top 10 (fast) → Cross-encoder reranker → top 4 (precise) You get speed + accuracy. Neither alone is enough. Using FlashRank - runs locally, completely free. Reranking isn’t optional if you care about answer quality. #BuildInPublic #RAG #AIEngineering #LLM

Google Bay View Part 2! 🏢 Using attractor points to lift the roof mesh. 📐 Master vertex manipulation with Cull Duplicates, Remap, and Graph Mapper for controlled Z-displacement. 🏗️ Pure computational logic! 🚀 #Rhino3D #Grasshopper #BIG #ComputationalDesign #ArchViz

It took less than 20 minutes to build this tool with the help of Gemini 🤯 As a non-technical guy, I was endlessly searching the internet for a specific image cropping tool. I needed something that could easily split photo grids, auto-detect objects, and let me draw custom shapes to crop out multiple items from a single image all at once. Everything I found was either too complicated, heavily restricted, or just didn't do what I needed. So, I decided to try building it myself using AI. I explained my vision to Google Gemini, and before I knew it, I had a fully functional web app running right in my browser! Introducing the Multi-Crop Tool ✂️ ✅ Grid Split: Perfectly slice 3x4 (or any size) photo grids instantly. ✅ Auto-Detect: Let the tool find and box objects for you. ✅ Custom Polygon: Draw precise freehand shapes around complex items. ✅ Batch Export: Download all your cropped images at once in a ZIP file. 🛠️ Try the tool for free here: https://lnkd.in/gxwa5Ds3 📺 Watch my quick tutorial on how to use it: https://lnkd.in/g2QSuGNx A huge thanks to Gemini for making this possible! It is incredibly empowering to go from an idea to a working product without needing to know how to code. The barrier to creating useful things has never been lower. 🚀 Have you built anything fun with AI recently? Let me know in the comments! 👇 #NoCode #ArtificialIntelligence #Gemini #ProductivityTools #ImageEditing #TechForEveryone #BuildInPublic #Innovation

🌍 I just shipped something I'm genuinely proud of. At Sylva, I've been building the Great Plains Earth Observation Platform — a cloud-native geospatial intelligence dashboard that turns raw Sentinel-2 satellite imagery into conservation intelligence for protected areas across Africa (Sapi, Mana Pools & Chewore). Here's the architecture that makes it run entirely serverless: 🛰 COG + TiTiler (Google Cloud Run) Sentinel-2 imagery is served as dynamic XYZ raster tiles with per-band rescaling — Red (0–0.7), Green (0–0.6), Blue (0–0.52). No pre-rendering. Every tile is computed on-demand. ⚡ Zarr + HTTP Range Requests (no backend) When a user clicks any point on the map, the browser translates lat/lng → array index and fires HTTP Range Requests directly to GCS. Raw NIR + Red band arrays come back and NDVI = (NIR−Red)/(NIR+Red) is computed client-side in milliseconds. 🧩 DuckDB WASM (serverless vector queries) GeoParquet files live in a GCS bucket. A DuckDB WASM sidecar queries them directly in the browser — bbox + zoom-level filtering means you only pull the geometry you can actually see. 🗺 Deck.gl + MapLibre (WebGL rendering) Real-time layer compositing, time-slider navigation, on-click pixel analytics, and 3D terrain — all in the browser. + RAG pipelines with Claude AI for LLM-powered conservation intelligence. Live demo: https://lnkd.in/gkuJapF6 #EarthObservation #Geospatial #CloudNative #WebGL #DeckGL #MapLibre #Zarr #TiTiler #ClaudeAI #RAG #Conservation #RemoteSensing #FullStackEngineering

🚀 Phase 3/8: Vectorless RAG is here Everyone talks about RAG with embeddings… but what if you skip vectors entirely? I’ve implemented Vectorless RAG inside the RAG Code 3 folder of my repo. Instead of embeddings, it uses structure, keywords, and LLM reasoning to retrieve better context. Why this is interesting: ⚡ No vector DB needed ⚡ More explainable retrieval ⚡ Works great for structured data This is part of my 8-phase deep dive into RAG systems. Curious how it works? 👇 https://lnkd.in/d3Re7W_8 #GenAI #RAG #AIEngineering #LLM

Geospatial Technologies essential keywords, daily Tips 🌎 : Keyword : Deck.gl: Visualizing Large Datasets Category :Web Mapping Deck.gl: Visualizing Large Datasets 🚀 Deck.gl is an open‑source WebGL‑powered framework developed by Uber that enables the rendering of massive geospatial datasets directly in the browser. By leveraging GPU acceleration, it can display millions of points, lines, and polygons with interactive zoom, pan, and filter capabilities, all while maintaining sub‑60 fps performance. The library’s component‑based architecture (React‑style declarative syntax) allows GIS engineers to compose complex visual layers—such as hexbin, arc, and 3D extruded polygons—without managing low‑level WebGL code, making it a go-to tool for high‑performance, scalable geospatial analytics and storytelling. #GIS #GeospatialAnalytics #DataVisualization #WebGL #DeckGL #OpenSource #Mapping #SpatialData #TechInnovation #DataScience #TechGeoMapping #EssentialKeywords

Anthropic built a system to hide leaks. Then leaked the system that hides leaks. 512K lines of Claude Code just got exposed via npm source maps. 1,906 files. 40+ tools. 108 hidden modules. Shane and I went through the entire codebase. Not headline summaries. The actual architecture. What we found: ✦ 46K-line query engine handling every LLM call ✦ Coordinator system spawning parallel worker agents ✦ KAIROS always-on mode with tick heartbeats ✦ Buddy pet system with 18 species and gacha mechanics ✦ Telemetry tracking your frustration signals ✦ 6+ remote killswitches via GrowthBook ✦ Animal codenames: Tengu, Fennec, Capybara, Numbat And the wildest part? "Undercover Mode", a system prompt telling Claude to hide that it's an AI. They built the hiding system. Then exposed the hiding system. Swipe through for the full architecture breakdown. Directory structures. Feature flags. Hidden features. Remote controls. All details verified directly from the npm source maps. What's your take on these hidden features? 🙏 Follow Alex and Shane for more builder-level breakdowns. ♻️ Repost if someone in your network needs to see what's actually inside Claude Code.

The algorithm making your RAG system fast isn't a transformer. It's a graph. Every time you build a RAG pipeline (FAISS, Pinecone, Weaviate), there's one algorithm quietly doing the heavy lifting: HNSW (Hierarchical Navigable Small World). Here's the intuition: Imagine you're looking for a friend in a massive city. You don't knock on every door. You first zoom out to the highway level, get close to the right neighbourhood, then drill down street by street. HNSW does exactly this — but for vectors in high-dimensional space. It builds a multi-layer graph: → Top layers = long-range "highway" connections (fast, coarse navigation) → Bottom layers = dense local connections (precise nearest neighbours) At query time, you enter from the top, greedily hop toward your target, and descend layer by layer until you land on the closest vectors. The result? Approximate Nearest Neighbour search in O(log n) — instead of brute-forcing all million embeddings. The trade-off is intentional: you sacrifice a tiny bit of recall for a massive gain in speed and memory efficiency. In production RAG, that's almost always the right bet. When I integrated vector search into my own project, watching HNSW retrieve the right chunks from 100K+ embeddings in under 10ms genuinely felt like magic — until I understood why it works. The secret sauce isn't just the LLM. It's the infrastructure around it. #LLM #VectorSearch #RAG #MachineLearning #AIEngineering