Assem Sohaib
Bensalah

Designed and trained deep reinforcement learning agents (PPO, TRPO, A2C) on large-scale simulation environments, with a published cascading-policy framework lifting task success from 67% to 84% across 100 evaluation scenarios. Built end-to-end MLOps tooling on PyTorch, Weights & Biases, and MLflow for experiment tracking, reproducible runs, and automated evaluation across hundreds of training jobs. Authored a model evaluation framework — measuring policy quality, adversarial robustness, and convergence stability — adopted as the lab's standard benchmarking pipeline. Mentored junior researchers and co-supervised two Master's projects.

Architecting an AI-native collaborative platform for scientific writing, built around agent-based human–AI co-authoring workflows. Designed tool-using agents and prompt-chained generation for drafting, revision, and refinement, supported by an integrated RAG pipeline using vector embeddings and semantic search over document context. Applied structured outputs and feedback mechanisms to improve determinism, citation grounding, and writing quality.

Architected a production LLM-driven conversational system for real-time language proficiency assessment, supporting adaptive multi-turn interactions. Designed prompt strategies and stateful conversation orchestration for contextual coherence across long dialogues, plus an evaluation framework for conversation quality, contextual accuracy, and engagement. Integrated ASR and TTS pipelines with streaming LLM inference for end-to-end voice interaction at near real-time latency.

Engineered a full-stack construction management platform end-to-end, from data model to client-facing UX. Designed a Supabase-backed schema with real-time data synchronization across concurrent on-site and office users. Built resource allocation and reporting modules to improve project transparency and stakeholder communication. Owned CI/CD, deployment, and observability on Next.js / Vercel.

Delivered an AI-powered virtual assistant SaaS automating a majority of routine business tasks for early customers. Built multi-agent workflows using CrewAI and custom Python services for orchestration, database retrieval, and report automation. Designed end-to-end UI/UX for chat, calendar, and analytics dashboards to improve task completion times.

Built an NLP spam detection system on top of a fine-tuned BERT model, deployed as a Google Workspace add-on. Owned data preparation, model fine-tuning, and integration into the production add-on while leading a small cross-functional team. Defined evaluation metrics and held-out test protocols to validate model behavior before release.