Launch
HowlHouse
Built an AI-only Werewolf spectator product with deterministic replay logs, a FastAPI backend, a Next.js viewer, recap/share artifacts, agent uploads, seasons, and tournaments.
London-based AI engineer
I build AI products, agent systems, and optimization-backed tools.
I’m Sohaib Benamor, a graduate AI / machine learning engineer with a Distinction MSc in Artificial Intelligence from King’s College London. I work on applied AI systems that survive contact with real users, not just demos.
My latest launch is HowlHouse, a spectator-first AI Werewolf product I built with Codex. I’m also building ContractFlow, an LLM-driven contract workflow tool focused on extraction, retrieval, and risk scoring.
Selected work
Projects that sit at the intersection of AI, product, and systems.
In progress
ContractFlow
An LLM-driven contract workflow tool that turns PDFs into structured extractions, clause retrieval, and policy-based risk scoring with offline evaluation in the loop.
Research & optimisation
TA allocation at King’s
My MSc dissertation automated large-scale teaching assistant allocation with MILP, candidate pruning, synthetic data generation, and rigorous constraint validation.
Latest writing
I built an AI Werewolf spectator sport with Codex, and the real work started after the first green build.
This write-up is the honest version of the project: Codex wrote the MVP quickly, but the real work was specification, hardening, CI, and taste.
- How I split the work between Codex and GPT-5.4 Pro
- What broke after the first green build
- Why security review, CI, and frontend identity mattered more than raw implementation speed
