Plainsight
Plainsight is the tooling we build for the AI era: assistants that work alongside our auditors today, and the guardrails and monitoring that agentic systems will need tomorrow.
The problem
Every team we talk to is shipping more code, faster, and understanding less of it. Review capacity did not grow with generation capacity, and agents now act on systems directly rather than proposing changes a person reads first.
The answer is not more output. It is knowing which claims about a system are actually checkable, and checking them. That is what this company has done since 2010, and Plainsight is that discipline pointed at the tools rather than at one engagement.
What Plainsight covers
Our auditors run every engagement with tooling that reads the codebase alongside them: surfacing the functions worth arguing about, drafting the preconditions each one assumes, and flagging where a caller does not honour them. It does not replace the reviewer. It makes sure the reviewer never runs out of leads before they run out of time.
Credentials leak into repositories constantly, and they leak faster now that agents write and commit code. Scanning for them is a solved problem in the easy cases and an unsolved one everywhere else: rotated keys, internal formats, secrets that are only secrets in context. We are building the part that needs judgement.
An agent with credentials is a piece of production software that nobody specified. Writing down what a system is allowed to do and then proving it stays inside that is the work we have done for two decades, and it is exactly what agents are missing. Guardrails, governance, and monitoring for agentic systems is where this is going.
How we build it
A finding is worth something when you can reproduce it. Everything the tooling reports carries the reasoning and the location that produced it, so an engineer can agree or disagree on the merits.
Models are very good at generating candidates and quite bad at knowing which ones are real. We use them for the first and put our own process, and our own people, in front of the second.
Preconditions, invariants, and specifications are not AI concepts, they are the vocabulary that makes an AI claim checkable. The tooling is built on the same formal methods the rest of our work stands on.
Plainsight is not a product you are left alone with. It is how our engineers work, and what stays behind when they are finished.
We build a model of the code and the behaviour it is supposed to have, because you cannot enforce a policy you have not written down.
Assistants generate candidate findings at a scale a team cannot; the process in front of them decides which ones survive.
Specifications become the guardrails and the monitors, so the assurance holds as the system and the agents acting on it keep changing.
Plainsight is early, and we would rather talk to teams with the problem than guess at it. If any of the above is on your desk right now, we want to hear how you are handling it.
See how we work
