Where's Waldo, for storage tanks. A VLM labels each tile as tank, part, or none. SAM 3 skips the none tiles (64%) and still gets 96.5% recall at IoU 0.5. A 4B gate ties the 32B F1 in half the time. thunderheadxpler.blogspot.co… #GIS #GenAI #ComputerVision

Plain-English SQL conditions, no API key. A Rust DuckDB extension runs Jev functions on a local Laya model. DuckDB does geometry; the model reads only text. 106 North Sea wells, zero model calls. thunderheadxpler.blogspot.co… #DuckDB #RustLang #GIS #GenAI

One photo + one sentence → 3D. Apollo Spatial VLM: a VLM points at what you ask for, a depth model adds meters, and the browser renders the scene as ~92k Gaussian splats. ~2 s per request on Blackwell, ~5 s on a Jetson AGX Orin. github.com/mraad/Apollo-Spat… #VLM #cuda @L42ARO

Not everything needs an MCP to be agentic. arcpi lets a pi agent call ArcGIS REST (the old way) and a live MCP map (the new way) from one code-mode script. Bonus: p(døøm) = 0, courtesy of OpenShell :-) thunderheadxpler.blogspot.co… #ArcGIS #MCP #GenAI #GIS #OpenShell

Presented ProCowork at Esri UC in San Diego: Claude Code in a chat panel inside ArcGIS Pro. Type a request; Claude writes the ArcPy, runs it on your live map, and shows the code and results. Experimental, YOLO mode, keep backups :-) thunderheadxpler.blogspot.co…

State in. Structured decision out. Measurable, so improvable. #Jev and #CLM share that contract. On my lunar lander, 8B CLM: 869 ms/decision. A small CLM trained in 42 s on a Mac: 19.8 ms, 96/96 matching guidance. It follows guidance; it doesn't fly solo. thunderheadxpler.blogspot.co…

I’m coining #GeoSI: fusing GIS and super-intelligence. agentic systems that understand place, learn through feedback, coordinate action, and improve outcomes. #GeoSI uses geography as foundation and feedback system. In this era of accelerating automation, we fight fire with fire!

The engineer and the trainee. Adaptive MPC knows the physics. Laya, a small decision model, learned from examples. Alone, Laya lands 0/90. With MPC vetoing about one in three of its moves, it gets 90/90, through a 60% engine loss. Fly it in your browser: github.com/mraad/lunar-mpc-l…

State + text in. Structured, measurable decisions out. I built Lunar Laya to try this on a Mac: 30/30 test landings and 58% less peak MLX memory with 6-bit weights. It follows explicit guidance requests. #laya thunderheadxpler.blogspot.co…

How small can a model be and still turn a question into a GIS query? GPT-2 medium + nanoGPT, unchanged training loop, 5 min on 2 GPUs. 89% exact match on North Sea wells/pipelines/discoveries. Bigger model: no gain. More data: +18 pts. thunderheadxpler.blogspot.co…

FELN #Qwen: Train on #RTX, Run on #Jetson #Orin thunderheadxpler.blogspot.co…

FELN WebGPU: A Language Model, a Database, and a Map in One Browser Tab thunderheadxpler.blogspot.co…

Thunderhead Explorer: FELN RAG: Five Examples and a Spatial Query thunderheadxpler.blogspot.co…

Thunderhead Explorer: FELN: Find Existing Location with N Layers thunderheadxpler.blogspot.co…

Thunderhead Explorer: Layers JSON: Giving LLMs the GIS Context They Need thunderheadxpler.blogspot.co…

Thunderhead Explorer: Nano Face: A Mac Camera, a Jetson, and One USB Cable thunderheadxpler.blogspot.co…

Tracking A Million Displaced People in Real Time: Lebanon’s Crisis Coordination System esri.com/about/newsroom/blog…

One Lander, Three Controllers: Reinforcement Learning, MPC, and Quadratic Programming - thunderheadxpler.blogspot.co…

Trusting Data: New Guardrails for the AI Era - esri.com/about/newsroom/publ… @Esri #GeoAI

Run #claudecode in #arcgispro as a coworker. github.com/mraad/ProCowork