For Students

We regularly collaborate with students on thesis projects at the intersection of AI, data systems, and real-world applications. Below are our open topics for spring 2027. Interested? Get in touch via our contact page.

Topic 1

Vectorless RAG Evaluation

Retrieval-Augmented Generation (RAG) systems are typically benchmarked using embedding similarity metrics, but these don't always correlate with real-world usefulness. This thesis would explore evaluation frameworks for RAG that operate without vector similarity as the primary signal — for example, using LLM-as-judge approaches, task-completion proxies, or faithfulness metrics. The goal is a reproducible evaluation methodology applicable to production knowledge systems.

Suitable for a student with interest in information retrieval, NLP, or applied ML evaluation methodology.

Topic 2

Embedding Evaluation Across Heterogeneous Data Types

Embedding models are commonly evaluated on text benchmarks, but production systems increasingly index structured data, code, and mixed-content documents. This thesis would systematically evaluate embedding model performance across these data types, propose a benchmark suite, and identify where current models fall short. Practical output: a selection guide for embedding models given a data profile.

Suitable for a student with interest in representation learning, NLP, or ML systems.

Topic 3

LLM Agents for Security Operations and GDPR-Compliant Threat Response

Security Operations Centers (SOCs) face growing alert volumes that exceed human analyst capacity. This thesis would investigate how LLM agents can be applied to SOC workflows — triage, investigation, and incident response — while operating within GDPR constraints around personal data handling and audit accountability. Key questions include: what tasks are safe to delegate to an agent, how to enforce data minimization in agentic pipelines, and how to maintain explainable, auditable decision trails. The goal is a practical model for deploying LLM agents in a security context where regulatory compliance is non-negotiable.

Suitable for a student with interest in AI systems, security operations, or data privacy law.

Topic 4

Inferring User Satisfaction from Conversational Signals in Enterprise AI Systems

Most enterprise AI deployments don't get useful feedback — users rarely click thumbs up or fill in ratings, so teams are largely flying blind on whether the system is actually helping. This thesis would look at whether you can infer satisfaction from how users behave in conversation instead: do they rephrase the same question? Drop off mid-thread? Say "thanks, that's exactly it"? The methodology would be built and validated on public datasets like LMSYS Chatbot Arena, WildChat, and Anthropic HH, then tested on how well it holds up in domain-specific enterprise Q&A settings.

Good fit for a student interested in NLP, user behavior, or applied ML.