Make world-model plus language-model judgment measurable.
Mohamed A M Elansary, PhD — multimodel hydroclimate forecast evaluation, geospatial and Earth-data hygiene, scientific/HPC rigor, and production agent evaluation sets for a physics-native multimodal stack.
Physics-grounded evaluation
- Six-plus years of multimodel, multi-basin surface-water and groundwater forecast experiments across hydroclimates on Linux/HPC.
- Compared statistical and physically based stacks, quantified uncertainty, and reported regime-dependent failure modes rather than a single flattering score.
- The posting is domain-agnostic on prior physics; atmospheric and hydrologic modeling is the sourced transfer, not an insurance catastrophe-model claim.
Geospatial data and agents
- USGS, NOAA, and NASA multi-source databases with publication-grade QA; ArcGIS/QGIS for eight years; Field GIS ArcPad 10.
- Production GPT, Claude, and Gemini agent workflows with retrieval, routing, tenant isolation, provenance, and regression evaluation sets at Vertexium.
- That maps to evaluating whether a tool-using trajectory reflects appropriate judgment and constraints. It is not 3D vision-language model training.
Proposed first contribution
For one world-model-plus-language-model workflow that already matters to underwriting, pricing, or mitigation, define intended judgment and a small failure taxonomy: physics disagreement, spatial mismatch, missing context, overconfident tool use. Stand up an evaluation set with provenance on traces, compare simple baselines, attach uncertainty and slice-level failure rates, and make regressions visible before wider adoption. This is a proposed measurement approach, not a claim of prior catastrophe-model or Stand product ownership.
Honest fit boundary
Insurance-specific catastrophe modeling and the Stand product stack are a stretch. I have not authored insurance or climate-risk publications, trained 3D vision-language models, or claimed Stand World Model ownership. I do not invent metrics, safety research, or RLHF. The credible contribution is physics-grounded evaluation, geospatial and Earth-data hygiene, scientific/HPC rigor, and production agent evaluation harnesses.
Role and location
Machine Learning Engineer - Multimodal Modeling · San Francisco · Ashby workplaceType Hybrid. The live posting body does not specify a hybrid day count. Willing to relocate to San Francisco with a relocation package. Remote eligibility is not asserted.
Posting compensation: “The annual base salary range for full-time employees in this position is $250,000 to $295,000 + meaningful Equity Grant.” Ashby display “$250K – $295K • Offers Equity”. · Official role posting