My role and contributions
I contributed to early parent-child chunking, RRF hybrid retrieval and NLI citation-verification design, then continued product support and iteration. CompliPilot and Attrax are the same product line.
Background
Cross-border sellers need to identify risks from product features and target markets, then check the regulations behind each conclusion.
Core principle
Trust in compliance AI is built on the citation layer, not the language model — in high-stakes settings, hallucination isn't a UX issue, it's a non-starter.
Current method (public repository, 2026-09-21)
Trace each claim to its regulation source
The current public workflow connects visual facts, applicability, source articles, generation and layered verification.
Product vision
Extract category, visible features and unresolved questions from product images as structured observations.
- Input
- Product image
- Output
- Category and visible features
Citation checks distinguish source / literal / semantic; unverified states remain explicit.
The current Attrax public repository describes a vision → generate → verify pipeline: identify product categories and visible characteristics, then combine mandatory checks from a rule-based knowledge base, regulation source text and model generation.
- Find mandatory regulations by market, category and product features, and load the source text.
- Preserve regulation references and source excerpts alongside generated findings.
- Distinguish source, literal and semantic verification states; missing verification remains unverified.
- Export compliance reports, profit analysis, decision tables and roadmaps.
These are implementation capabilities documented by the public repository. This case study does not provide production-outcome or performance validation for that version.
Historical design and evolution
The earlier design used parent-child chunking, RRF hybrid retrieval and NLI citation verification, with Must-Check and embedding fallbacks. The steps below describe that period. The current public implementation centers on the rule-based knowledge base and regulation source text, without an embedding dependency.
Parent-child chunking
Balances recall completeness with retrieval precision.
RRF hybrid retrieval
Fuses multiple retrieval passes to lift recall of relevant sources.
NLI hard gate
Every claim must be supported by a retrieved source — otherwise it does not pass.
Must-Check + fallback
Critical items get mandatory verification, with embedding fallback for robustness.
Why it matters
The project established a way to present regulation sources, generated conclusions and verification status separately, enabling users to inspect the evidence and identify content that still needs human verification.
Lessons learned
As the design evolves, update data sources, citation contracts and verification boundaries together, and preserve real evidence for each version. Describe implemented functionality, available citations and verified conclusions separately.