Waseem Alshikh

Waseem Alshikh is the co-founder and chief technology officer of Writer, where he develops specialized language models, financial AI evaluations, and autonomous agents for enterprise work. His central concern is whether these systems can act reliably when documents are incomplete, evidence is missing, or an apparently helpful answer would be dangerously wrong.

Alshikh grew up south of Damascus and studied computer science in Lebanon. While learning English, he built software to condense academic material, replacing an inadequate rules-based approach with statistical methods. Graduate study in Jordan and work in Dubai preceded his partnership with May Habib, which began in 2013 with enterprise translation and localization through Qordoba. They founded Writer in 2020, with Habib as chief executive and Alshikh leading technology.

At Writer, Alshikh helped develop the Palmyra model family, including models specialized for finance and healthcare. His work treats domain knowledge, evidence handling, operating cost, and safe interaction with enterprise systems as separate engineering requirements, not qualities guaranteed by impressive general-purpose benchmark scores.

• FailSafeQA and context grounding. Alshikh co-authored FailSafeQA, a financial evaluation that tests malformed or out-of-domain questions alongside missing, irrelevant, and OCR-corrupted documents. The public dataset separates answer robustness from context grounding and appropriate refusal. His AI Engineer Summit presentation demonstrates why reasoning-oriented models can appear capable while producing answers unsupported by the supplied evidence.

• Governed enterprise autonomy. In his engineering work on Action Agent, Alshikh describes software that executes multistep tasks using an isolated computing environment, browser and terminal access, readable plans, and iterative correction. Sandboxing, permissions, auditability, and human supervision are integral to deploying these systems inside organizations.

• When intervention makes agents worse. Alshikh co-authored research on agent failure prediction showing that even accurate critics can reduce overall performance when they interrupt trajectories that would otherwise succeed. The practical implication is to evaluate interventions against the specific task distribution before assuming additional oversight improves outcomes.

• Efficient tool-using models. The Palmyra X6 technical report, which Alshikh co-authored, describes an enterprise-oriented model trained on verified synthetic tool-use trajectories through anchored supervised fine-tuning. His more recent work extends Writer’s focus from reliable answers to longer-running autonomous workflows with explicit governance, operational safeguards, and cost controls.