OnSiteSafety: how a $5K class became a productized AI stack
OnSiteSafety is the construction safety training brand we launched in 2024 to test a simple thesis: that an AI infrastructure company should operate at least one productized service end to end before it asks clients to do the same. Two years and a hundred plus paid classes later, the architecture is stable enough to write down.
The product is short form construction safety training. Five to nine thousand dollars per class, delivered to a client crew, with the entire content pipeline operated by AI infrastructure we built. A typical class includes a regulatory review specific to the trade, scenario based training material adapted to the client’s job sites, comprehension testing, and post training documentation. None of this is generated on the fly. Every class is curated, reviewed, and signed off by a human safety officer before delivery. The AI infrastructure shortens the production cycle from days to hours.
The architecture has four layers. At the bottom, a content corpus of WorkSafeBC regulations, OHS code, and our own training material, indexed in a vector store. Above that, a retrieval pipeline that pulls relevant regulations and case law for any given trade and job site combination. Above that, a generation layer that drafts training material in our standard structure and tone. At the top, a review layer where a human safety officer reads, edits, and signs off. Sounds straightforward. The hard part is the operational discipline.
The lessons we have carried into client engagements are concrete. First, productize the parts that repeat and keep the parts that vary human. We did not try to automate the safety officer; we automated the slide assembly, regulation lookup, and rough draft. Second, monitor the outputs ruthlessly. We score every class on a rubric and review the bottom decile every week. Third, do not undersell the human in the loop. Clients pay for the safety officer’s judgment as much as for the training material. The AI is the production tool, not the product.
The reason we tell this story is that it answers the question every prospective client asks: have we operated AI infrastructure at production scale ourselves. The answer is yes, for two years, with paying customers, on a system we built. When we recommend an architecture to a client, we are recommending it from operator experience, not from the documentation.