Selected work

Applied AI · Knowledge retrieval · Monitoring

Compliance AI

Making complex regulatory work easier.

Corporate teams in enterprise environments perform a range of time-consuming tasks with limited resources. Our task was to find a solution that would simplify these tasks for both corporate teams and internal consultants. Examples include monitoring legislative changes, checking contracts and methodologies for compliance with regulatory requirements, answering repetitive questions, and looking up details in contracts, methodologies and rules. AI analysis and retrieval-augmented generation (RAG) systems provide tools to simplify this work.

At a glance

End-to-end product ownership: turning business requirements into connected workflows for monitoring, research and document review.

In use and in active development. No formal answer-quality evaluation completed.

Contribution
Stakeholder analysis, UX, architecture & implementation
Period
May–September 2026 · Ongoing
Status
In use by agency consultants · Active development

Compliance AI is a RAG-based knowledge base with a chatbot-style interface. It answers regulatory questions using only approved sources, manages a custom database of approved source documents, and analyses uploaded documents for compliance with existing regulations.

01

The challenge

Agency consultants need to follow changing Czech and EU regulations, find relevant knowledge and review internal policies. The product brings these related tasks into one application.

The UX challenge was to make a complex information workflow understandable: what changed, which sources support an answer and what still needs human judgement. A fluent AI response is not enough when the underlying evidence matters.

02

My role

I owned the work from stakeholder analysis to the working application. I gathered needs, researched competing products and wrote user stories to translate business requirements into product scope.

I designed the frontend and application architecture, then implemented the product with AI assistance. My design decisions covered the boundaries between research modes, access to supporting sources and behaviour when evidence is missing.

I also created implementation tests and a separate answer-quality evaluation framework. Formal answer-quality evaluation has not yet been completed.

03

The decisions

Organise the product around distinct research tasks

Monitoring brings relevant material into the knowledge base. Stored-knowledge questions and wider online research have separate modes; document review extends the product to internal policies and potential regulatory risks.

This gives each task a defined place in the product and makes the source boundary explicit when assessing an answer.

Make the evidence part of the experience

Knowledge-base answers respect access rights and check quotations against stored sources. Reviewers can inspect the material behind a response.

The design supports a person checking the evidence. A matching quotation alone does not establish that the interpretation is correct.

Treat missing evidence as a product state

The knowledge-base mode can decline to answer when supporting evidence is unavailable.

Refusal makes an information gap visible. It preserves a place for further research and human judgement rather than presenting an unsupported answer as resolved.

Evaluate the answer separately from the software

I separated implementation tests from a framework for evaluating retrieval and answer quality.

A deployed feature and an answer people can rely on need different evidence. Current use is confirmed; answer reliability remains to be formally evaluated.

FastAPIReactSupabaseHybrid retrievalQdrantOpenSearch
04

The result

The agency uses the AI knowledge base and legislative and regulatory monitoring in client-project work. I delivered the product from requirements through design and implementation; the recording above demonstrates the monitoring module.

The wider application also supports online research and policy review. Development continues on scanned-PDF support, retrieval ranking and source coverage. No formal answer-quality evaluation has been completed, and no measured accuracy or time-saving result is claimed.

What I’d measure next
Measure
Task completion, time to verify an answer, source support and appropriate refusal.
How
Test representative consultant tasks: identify a relevant change, find supporting evidence and assess an answer. Observe completion, mode confusion and verification time. Separately, use expert-reviewed questions, including unanswerable cases, to score correctness and whether sources support each claim.
Check alongside it
This is a proposed validation plan, not a completed study. Check access-rights failures, unsupported claims, missed refusals and source coverage separately.

Explore more work.

11 projects

Scroll to explore all projects.

Pavol JuríkProduct / UX / AI

Let’s talk

Good things start
with a conversation.

Have a role, a project, or an idea in mind?
Tell me a little about it.