Salman Parvez | ML Systems

Quiet hours after a human change is a smart one. On a job site the next trade doesn't start until the last change is checked. Refusing any guardrail softer than the client's policy is the right default, and the same rule we use at ML Systems.

33% to 61% with the same model is a great example. Every gain came from the rules and checks around it, the way a crew gets better with the same people and a better process. The safety gate hitting 100% is the number I'd watch first.

Keeping the math as plain functions separate from the UI is the part I like. An estimator can check the formula itself. I'd love to see waste factors as an input on drywall and flooring. We'd use something like that at ML Systems.

Pointing the approver at the risky files first is how a good inspector walks a job. Nobody checks every nail. They go straight to the connections and anything carrying load.

The summary that drops the condition is the one I'd worry about most. "Exchange within a week if the packaging is sealed" turning into "exchange within a week" is exactly how a spec note turns into a change order. A source, a last-checked date, and an owner on every item is the standard we're holding ourselves to at ML Systems.

Writing DROPPED next to the rejected option is the same reason superseded drawings get a big VOID stamp instead of a note in the meeting minutes. If the old sheet stays in the set without the stamp, someone builds from it. "Kept minimal for later" in the summary is a great tell to watch for.

Turning every escalation into a concrete test is how good contractors handle callbacks. You can't predict every one, but you can make sure the same problem never brings you back twice. That's a habit we're building into ML Systems from the start.

Tiering by risk instead of by author matches how inspections work on a job site. The electrical rough-in gets inspected whether a veteran or an apprentice pulled the wire, and nothing gets closed up until someone signs off. Reviewing the spec and the evidence instead of every line is how a good inspector spends their time too.

Coming from estimating, I'd add the schedule of values. Pay apps usually bill percent complete line by line, so checking each line's billed percent against its inspection record catches overbilling well before the total balance does. Turning "the work is complete" into a recorded inspection is the same instinct we're building around at ML Systems.

An empty result set carrying no signal about whether the store is empty or the bar was too high is a sharp catch. I'd want the caller to log which case it was, so nobody reads silence as "no relevant memory." Testing it on a solo contractor's working rules was a nice touch too.