About

ReliaParse is a research project at the University of Helsinki on reliable, auditable LLM-assisted analytics. We work with a small number of pilot organisations to develop and validate methods that make large-language-model-driven data work trustworthy enough to use on real decisions.

What we do

Analytical teams in finance, public administration, life sciences and other regulated settings are under pressure to adopt LLMs. The everyday reality is brittle: a chat answer that looks plausible, a one-shot script that nobody can re-run, results that cannot be defended in a review. We are building the missing layer — methods that combine LLMs with deterministic tools so that every generated artefact comes with a clear chain from intent to result.

The work has three strands that feed each other:

• Intent verification. Methods that check whether a generated implementation matches a stated analytical intent, before it touches a stakeholder.
• Workflow integration. Field studies of how analytical teams actually work today, and where reliable LLM assistance fits without breaking the rest of the pipeline.
• Pilot engagements. Scoped collaborations with real organisations on real workflows, run under written agreements that protect both sides.

How we work

We are a small group inside the university, with backgrounds in machine learning, software engineering, applied statistics and tech transfer. We publish what we learn, and we are explicit about what is still unproven. The blog on this site is part of that: short notes from the field, written in the same voice we use with pilot partners.

If your team has an analytical workflow where correctness matters and you are curious whether reliable LLM assistance could fit, we would like to hear about it. A short note describing the workflow is the right place to start.