ask a general-purpose AI a work health and safety question and it will answer with total confidence. sometimes it is right. the problem is the failure modes, because they are exactly the ones a non-expert never catches: section numbers that do not exist, NSW regulations quoted in a western australian context where the law is genuinely different, a generic hierarchy-of-controls template where an ICAM investigation was needed.
in a safety-critical domain, a confident wrong answer is worse than no answer. that is the problem safetylm exists to solve.
what it is
safetylm is an open-source AI assistant for AU/NZ WHS practice. instead of trusting a model’s memory, it retrieves the relevant primary sources at query time (legislation, regulations, codes of practice), reasons over what is actually in front of it, and cites what it used. jurisdiction is a first-class filter, not an afterthought. and it runs entirely locally: no API dependency, no per-query cost, full data sovereignty.
when it cannot find a grounded source, it says so. that honesty is the product.
what building in public means here
building in public is not a marketing posture. concretely, it means:
- the roadmap, architecture and corpus criteria are published in the repo before the results exist.
- each phase has explicit acceptance criteria, so “done” is checkable by anyone.
- the changelog records what actually shipped, not what was hoped.
it also means the current status is public: the corpus build is in progress, with 616 AU/NZ source documents catalogued and processed into 20,451 metadata-tagged chunks. the interesting parts are still ahead.
the benchmark matters more than the model
the deliverable we care most about is safetylm-eval: a benchmark of 500+ validated WHS questions with ground-truth answers, published under cc by 4.0. without a domain benchmark, WHS AI quality is unmeasurable, and every vendor claim is just vibes. a shared yardstick is valuable regardless of how safetylm itself scores against it.
if you are a WHS practitioner, your domain judgement is the most valuable contribution the project can get. the best way in is an issue on the github repo.