¸£ÀûÔÚÏßÃâ·Ñ

DSP Seminar: LLMs That Lie Less: Engineering Responsible Language Models for Production

Data Science Platform Seminar Series VIII

Talk Abstract:

Large Language Models are increasingly being deployed in high-stakes applications, yet most discourse around their risks remains at the policy and philosophy level. This talk takes a different approach: responsible AI as an engineering discipline.

Drawing on hands-on experience deploying LLMs in a production EdTech environment, this talk walks through the concrete techniques that separate a convincing demo from a system you'd actually trust with real users. We cover retrieval-augmented generation and the subtler problem of source adherence by getting models not just to access correct information, but to actually use it. We examine guardrail architectures, their latency and accuracy tradeoffs, and the surprisingly non-trivial question of who decides what a "safe" output looks like in context. We look at evaluation strategies that go beyond standard benchmarks and actually measure what matters: factual consistency, grounding, and failure modes that only surface in the wild. Finally, we discuss cost-reliability tradeoffs: How mixture-of-experts architectures and knowledge distillation let you build systems that are both responsible and economically viable.

This is the eighth talk in the 2025/6 Data Science Platform Seminar Series.

.

Speaker’s Bio:

Natalia Mallia is a Malta-based AI engineer and thought leader specialising in ethical, production-grade artificial intelligence.

Her work spans the design and deployment of LLM systems across EdTech, manufacturing, and creative industries, with a focus on responsible AI, production reliability, and cost-efficient architectures. She has spoken at NVIDIA GTC, SiGMA, and the World AI Cannes Festival, and was recognised at Malta's Best Businesswoman Awards in 2025.