Riyadh, Saudi Arabia
I design and build data-first AI systems — taking raw operational data, structuring it, and putting it behind LLMs so teams actually get intelligent features, not just dashboards. I work mainly in Python + FastAPI + OpenAI with RAG (retrieval-augmented generation), vector databases (Pinecone/FAISS), and prompt/guardrail patterns to power things like seller assistants, catalog enrichment, and internal policy chatbots for e-commerce-style platforms. Underneath that AI layer, I still have the foundation of a Fabric Certified Data Engineer — I know how to make data land cleanly in Microsoft Fabric, model it, and move it through ETL/ELT (Data Factory, Synapse/Dataflows) so the LLM isn’t guessing. I’ve worked with SQL Server, Oracle, and MySQL and I’m comfortable turning those sources into analytics- and AI-ready tables. What I enjoy most is connecting the two stacks — data → embeddings → retrieval → generation → BI/API — so a business stakeholder can open Power BI or an internal tool and actually use an AI feature that’s grounded in their own documents and policies. I’m currently focusing on LLM application patterns for marketplaces/e-commerce (policy-aware generation, support deflection, content cleanup) because they transfer directly to platforms.
● Built an AI-powered Saudi market-entry assistant for foreign founders, GMs, legal teams, expansion leads and consultants seeking clearer guidance on entering and operating in KSA. ● Designed the product to ingest, chunk, embed and retrieve official Saudi source documents, especially MISA and ZATCA material, so answers are grounded with citations rather than generic AI output. ● Mapped business activities against SSIC/ISIC-style classifications and flagged foreign ownership, restricted/regulated activity issues and special-handling questions. ● Generated practical outputs including licence summaries, approval sequences, compliance checklists, risk notes, regulator notes, law-firm questions and exportable PDF/HTML-style reports. ● Technical stack: Next.js, Supabase, PostgreSQL/pgvector and AI/RAG architecture with retrieval, source attribution and structured report generation
Built the Continuous Improvement Dashboard comparing baseline vs. actual milestone durations (“Approval to Develop” → “Approval to Deliver”) with department-level rollups. Bridged legacy SAP and S/4HANA with interim data models and standardized milestone definitions to maintain comparable trends during migration. Developed KPI and audit/governance dashboards with drill-through to surface slippage, compliance gaps, and phase-gate bottlenecks. Introduced lightweight requirements templates and feedback loops to stabilize release cadence and stakeholder satisfaction.
Built an LLM Policy Reply Assistant (Python, FastAPI, RAG over policy docs/macros, Pinecone/FAISS) that raised first-pass acceptable appeals +25–32%, cut repeat contacts −12–14%, and met P95 <2s latency. Shipped a Return/Chargeback Evidence Pack Generator (LLM + vision OCR, reason-code templates) improving win-rate +8–12% and reducing agent handle time −20–25%. Implemented ingestion→chunk (400–700 tokens)→embed→retrieve pipelines with freshness checks and routing by violation/reason code to keep outputs auditable with line-level citations. Added safety/latency guardrails (decision-language filters, ID/policy validators, caching) to meet production SLAs and enable help-center embedding.