Mohamed A M Elansary, PhD
Target: Research Engineer, Knowledge Team — Anthropic
Sourced insights
- Contextual Retrieval: Anthropic shows traditional RAG often loses document context in chunks; contextual embeddings + contextual BM25 cut top-20 retrieval failure by 49%, and with reranking by 67% (5.7% → 1.9%) — information architecture designed for LLM consumers, measured with hard retrieval failure rates. Source: anthropic.com/engineering/contextual-retrieval
- Dynamic filtering for agentic web search: Claude's web search/fetch tools write and execute code to filter results before they enter the context window — ~11% average accuracy lift on BrowseComp / DeepsearchQA with ~24% fewer input tokens (e.g. Opus 4.6 BrowseComp 45.3% → 61.6%). External knowledge is filtered for the model, not dumped into it. Source: claude.com/blog/improved-web-search-with-dynamic-filtering
Proof — shipped work + PhD UQ
- Production multi-tenant agentic LLM / RAG-adjacent systems (Claude, GPT, Gemini): retrieval, query routing, per-tenant data isolation — WhatsApp AI receptionist + voice booking agents in live use (Vertexium).
- Reliable Python/API automation for compliance, CRM data quality, and environmental monitoring workflows — shipping over demos.
- PhD Environmental Engineering, TAMUK 2022: hydrologic uncertainty quantification — multi-basin, multi-hydroclimate ensemble forecasts on HPC (MODFLOW, VIC, PIHM, NASA LIS; USGS/NOAA/NASA data) as failure-mode / hard-eval rigor for knowledge systems.
- Honest frame: applied knowledge systems + shipping — not pure interpretability or foundation-model pretraining.
- The PhD who ships. Prefer take-home / work-sample when interview format allows.