back to all research
building in publichealth & safety
safetylm
an open-source AI assistant for AU/NZ WHS practice, built in public.
why this matters
general-purpose AI fails WHS practitioners in ways a non-expert never catches: hallucinated section numbers, NSW regulations quoted in a WA context, generic advice where an ICAM investigation was needed. in a safety-critical domain, a confident wrong answer erodes trust and can contribute to poor decisions.
what we're doing
- retrieval-augmented generation over primary sources only
- jurisdiction as a first-class filter
- conservative confidence: saying "i couldn't find a source" is a feature
- safetylm-eval, a 500+ question benchmark with ground-truth answers, published cc by 4.0 so any WHS AI can be measured
- open methodology end to end
where it's at
corpus build in progress (phase 1 of 6). the roadmap, changelog and acceptance criteria are public in the repo.