Hogeschool PXL · GÉANT · Belnet

EXPATS

Explainable Autofill for Trustworthy Surveys

Two AI co-pilots that respect your data: one helps you design better questionnaires, the other helps respondents answer them with cited evidence from their own documents.

Built by researchers. Made for institutions. Developed in the EU.

What it is

One platform, two co-pilots

EXPATS pairs a survey response assistant with a survey design assistant. Both run on the same privacy-first foundation, and both explain their reasoning instead of asking you to trust a black box.

Cue for respondents

Answer with evidence, not guesswork. Cue reads documents you upload, retrieves the relevant passages, drafts a concise answer, and cites exactly where each part came from.

  • Retrieval-augmented drafting from your own files
  • Every suggestion carries its sources & highlights
  • PDF, Word, PowerPoint, Excel, text, Markdown & images — processed, then discarded

Shape for survey authors

Design better questionnaires, faster. Shape suggests and rewords questions, checks them against style and consistency rules, proposes tags, and imports or exports surveys in standard formats.

  • LLM suggestions, rewording & auto-tagging
  • Deterministic style & consistency checks
  • QTI 3.0 plus LimeSurvey, Qualtrics & SurveyMonkey adapters

What it does

Project scope

EXPATS is a focused proof-of-concept. Being transparent about its limitations is part of being trustworthy, so here is the line, in both directions.

What it can do

  • Import and export surveys in LimeSurvey, SurveyMonkey, Qualtrics and QTI 3.0 format
  • Suggest, reword, validate and tag questions with an LLM
  • Draft cited answers from uploaded documents
  • Run fully on-premise with a bundled Keycloak login (OIDC)
  • Isolate every session and clean data up automatically (TTL)
  • Show exactly which source informed each suggestion

What it doesn't do

  • Audio / video transcription of documents planned
  • Conditional / skip logic in questionnaires planned
  • Matrix & grid questions (imported as single choice) planned
  • Local / self-hosted LLMs — cloud via OpenRouter today planned
  • Host or distribute surveys — it complements your tools, doesn't replace them by design
  • Profile users or track them across sessions by design
  • Fine-tune or train on your data by design

Design principles

Why it's built this way

A few key decisions have shaped this project. We trade flashiness for trust, portability and a code base that remains simple enough to audit.

Privacy by default

Session-based isolation and automatic TTL cleanup. No profiling, no cross-session tracking — privacy is the default, not a setting.

Explainable by design

Every suggestion carries its citations, and each session produces an audit report. You can always see why the system proposed what it did.

Interoperable

Four available platform adapters and an SDK-first REST API, so EXPATS slots into existing survey tools instead of demanding new ones.

Self-hostable

One Docker Compose stack, bundled Keycloak for login, and bring-your-own LLM key. Run the whole thing inside your own walls.

Roadmap

Where the project stands

EXPATS runs a pilot from January to May 2026 with Hogeschool PXL and partner institutions, demos at GÉANT TNC in June 2026, and reports out by the end of June 2026.

  1. Done

    Foundation & core modules

    Shared infrastructure, document processing & vector search, the Cue and Shape assistants, and a browser-based UI for each — all built and tested.

  2. Now

    Integration & pilot

    Deployment to pilot partners, real-user testing, and collecting metrics on authoring time, response time and citation accuracy.

  3. Next

    Demo & dissemination

    Live demo at GÉANT TNC (June 2026), final evaluation report (June 2026) and the open-source release for the GÉANT community.

On the horizon

Specced and waiting, to revisit after pilot feedback:

  • Conditional / branching logic
  • Matrix & grid questions
  • Local LLM support
  • Pilot metrics dashboards (MLflow)
  • Multilingual UI