Oussama Hafferssas is an AI and machine-learning engineer whose work at Trainline applies computer vision, mobile-platform engineering and production software practices to rail travel. His work on Trainline’s Travel Assistant addresses challenges beyond conversation: helping passengers navigate disruptions, find alternative trains, request refunds and reach human support when necessary.
His technical trajectory began with computer vision: in 2015, he co-authored a master’s thesis on video surveillance and adaptive Gaussian mixture models. He subsequently worked on mobile applications, including banking software at Crédit Agricole, before moving into Android platform architecture and developer productivity. His technical writing ranges from debugging Android App Bundles to explaining Jetpack Compose through Kotlin fundamentals.
By mid-2024, Hafferssas was a staff engineer on an Android-focused platform team and had begun experimenting with local language models, fine-tuning and a code-review hackathon project combining graph-based retrieval with LangChain. He argued that working with language models should become a basic software-engineering capability, alongside practical skills in prompting, model integration and retrieval-augmented generation. His analysis of AI-assisted development anticipated greater emphasis on reviewing generated code, orchestrating tools and maintaining the foundational knowledge needed when automation fails.
- Production agents combine deterministic and probabilistic systems. Conventional software checks remain necessary, but model-driven behavior also demands offline and online evaluation of accuracy, customer-facing tone and practical usefulness.
- Rail travel makes agent reliability a concrete operational problem. Different carriers, ticket conditions, delays and refund rules force assistants to resolve real constraints; Trainline also uses predictive machine learning to anticipate disruptions.
- Observability must work across the organization. Tracing tool calls and agent behavior supports failure diagnosis while giving product teams and other nontechnical colleagues access to operational insights without depending on engineers to retrieve every answer.
These priorities shaped his joint Trainline and Braintrust workshop on shipping production AI applications, where he described the demands of deploying customer-facing travel assistance at scale.
