Assem Sohaib
Bensalah

Ziritex is a collaborative scientific writing platform where I design agents that help researchers draft, revise, and refine documents. Its retrieval pipeline uses document context, vector embeddings, and semantic search to ground the writing. I use structured outputs and feedback to improve consistency and citation quality.

At RELA(CS)², I researched reinforcement learning for power-grid resilience. I trained PPO, TRPO, and A2C agents and evaluated a published cascading-policy framework across 100 simulation scenarios. I also built experiment tracking and evaluation tools with PyTorch, Weights & Biases, and MLflow.

For Verbly, I built an LLM-powered language practice system that adapts across multiple conversation turns. I designed the conversation state, prompts, and evaluations for quality and contextual accuracy. Speech recognition, speech synthesis, and streaming inference support its voice conversations.

+3

For Rochemère Systems, I built a construction management platform with Next.js, Supabase, and Vercel. It synchronizes project data between on-site and office users and includes resource allocation and reporting modules. I owned the data model, user experience, deployment, and observability.

For AutopilotVA, I built virtual assistant workflows that automated routine tasks for early customers. CrewAI and Python services coordinated agents, retrieved database records, and generated reports. I also designed the chat, calendar, and analytics interfaces.

At Metrolabs, I built a spam detector using a fine-tuned BERT model and integrated it into a Google Workspace add-on. I led data preparation and model training, then defined evaluation metrics and held-out tests before release.