Shubham Shukla

A hardware security module is a box whose entire job is to refuse. It holds keys it will never hand back, signs what you ask, and logs the attempt either way. Nothing about the interface is convenient, and that is the design.

Setting up a security world means deciding, up front and in writing, who can do what — how many cards it takes to unlock, who holds them, what happens when one is lost. There is no admin override. Systems that behave this way are rare, and they are the only ones I have never seen quietly bypassed under deadline pressure.

I have carried that shape into product work since. Write down the constraint before the feature. Make the recovery path explicit rather than implied. Prefer a system that says no clearly over one that says yes and hopes.

Replace this text with your own version of the post. The structure here is what a finished piece looks like in the panel: a lead paragraph, three or four sections, and a closing line worth quoting.

I joined Sparrow Charts as a senior engineer and left as CTO. Between those two lines was an engineering manager role that lasted eleven months and taught me more than either of the titles around it.

The work that stayed the same: reading code carefully, writing down decisions, and shipping something small every week. The work that changed entirely: my output became other people's output, which meant the feedback loop stretched from hours to quarters.

The unwarned part is how much of the job is repetition. The same architectural principle explained in five different rooms until it becomes the default. That is not a communication failure — it is the mechanism.

Replace this text with your own version of the post.

The naive version of this problem is a CSV importer per platform. The real version is deciding what a transaction is, once, in a way that survives five vendors disagreeing about fees, timestamps, and what counts as a transfer versus a trade.

We settled on a normalised internal event and a per-platform adapter that could only emit those events. Anything an adapter could not express became an explicit unsupported case rather than a silent guess. Silent guesses are what make tax software untrustworthy.

The lesson that generalises: when data comes from sources you do not control, spend your budget on the canonical model, not the importers. The importers will be rewritten. The model should not be.

Replace this text with your own version of the post.

Three advisory seats sound like three standing calls. They are not, and if they become that, nobody is served well. Founders need a fast answer to a specific question, usually within a day, usually in writing.

So the arrangement is asynchronous by default: send the question with the context attached, get a written answer with the reasoning shown. A call happens when the decision is genuinely two-sided, not when it is a status update.

The measure of whether this works is simple. If a team can make the next decision without me, the advice landed. If they are waiting on me, I have become the problem I was hired to remove.

Replace this text with your own version of the post.

At five engineers, each person is a fifth of your engineering culture. That maths is unforgiving, and it means the first filter is not seniority — it is whether someone writes down what they decided and why.

The interview that predicted the most for us was small, real, and paid: a scoped change against a repo they had never seen, with a written explanation attached. Whiteboard puzzles predicted nothing.

The part that took longest to learn was rejecting well. A candidate who gets a clear reason within two days tells other candidates about you. That is most of early recruiting.

Replace this text with your own version of the post.