Ian kiprono

Father & 🇨🇦Sr Engineer @pinterest | MSc Stanford | Building to $1M ARR. 🟩🟨⬜️⬜️⬜️

[Ian kiprono @Carson_001]

If you cannot endure until you see the results you want, then you probably shouldn't even get started.

[Ian kiprono @Carson_001]

@abadojack The tax here is based on bracket tax system which often sounds punitive for high income. The good thing is, it’s easier to risk starting businesses and companies as you can write the losses off from the taxes they would have gotten from you.

[Ian kiprono @Carson_001]

I officially became a Canadian Permanent Resident.

A few years ago, I arrived with a suitcase full of ambition and a lot of uncertainty about what the future would look like.

Since then, I’ve built products, grown my career, started companies, become a father, and met incredible people along the way.

Grateful for the opportunity, the lessons, and everyone who helped make this possible.

The journey continues.

Let’s build.

[Ian kiprono @Carson_001]

The agent that costs me $200/month handles 90% of support tickets correctly.

The one that costs $20/month handles 60% correctly.

Guess which one I'm keeping?

The $20 one.

I handle the other 40% myself.

[Ian kiprono @Carson_001]

It surprises me how rarely people talk about agent fatigue.

Not the model's.

Yours.

Running agents means constant evaluation. Constant judgment calls about what's good enough. Constant monitoring of systems that can silently drift.

The people who make this look effortless are either not doing it seriously or not being honest about what it actually takes.

[Ian kiprono @Carson_001]

I find it strange that I know which Narrareach users will churn before they do.

The signal is in session length.

Users who log in and stay for more than eight minutes in their first week almost never leave.

Users who open it, look at one screen, and close it in under three minutes — I've lost almost all of them by week four.

I have this data.

I still don't know what to do with it in time to change the outcome.

[Ian kiprono @Carson_001]

Something I find genuinely counterintuitive about agents on payroll:

The more you use them, the better you get at tasks that don't involve them.

Writing a precise brief makes you more precise.

Specifying exactly what you want forces you to actually know what you want.

The discipline of working with agents has made me better at working without them.

[Ian kiprono @Carson_001]

It amazes me that a competitor I respected built a comparison page designed to make Narrareach look bad.

I find it more revealing than damaging.

A comparison page tells you what someone is afraid of.

I read it twice and came away with a clearer picture of our positioning than anything I'd written myself.

Useful feedback. Uncomfortable source.

[Ian kiprono @Carson_001]

I find it genuinely amusing that the most valuable agent workflow I've built this year was something I didn't plan.

Set up one agent for a specific task.

Finished fast. Time left over.

Gave it an adjacent task. It handled that too.

The best systems I've built have emerged.

The ones I planned in detail rarely survive contact with actual use.

[Ian kiprono @Carson_001]

My two-year-old asked what I do for work.

"I teach computers to talk to people."

"Do they listen?"

"Sometimes."

"Like me?"

"Exactly like you."

Turns out parenting and AI development have more in common than I thought.

[Ian kiprono @Carson_001]

Something I didn't expect about reaching profitability at Narrareach:

The questions changed.

Before: will this feature help us survive?

After: will this feature compound over the next year?

Same features. Different questions.

The second one is harder to answer but leads to much better decisions.

[Ian kiprono @Carson_001]

I find the idea of fully autonomous agents genuinely misleading.

The best-running agents in my stack require constant attention — just different attention than writing code does.

You're not supervising the execution.

You're supervising the outcomes.

That's a full-time job in its own right.

The word autonomous implies you've stepped back.

You haven't.

[Ian kiprono @Carson_001]

I find it interesting that the feature I was most reluctant to build has become the one users mention most in conversations.

Subscriber attribution.

I thought it was too niche.

I thought writers wouldn't care about that level of detail.

I was wrong about both.

The features that feel obvious to build are often not the ones that actually change behavior.

[Ian kiprono @Carson_001]

It surprises me how often agents surface insights about my product that I'd been sitting next to for months without seeing.

Fresh context beats familiarity.

An agent reading the same data I look at every day will sometimes find a pattern I've been too close to notice.

That's not intelligence.

That's just what happens when something looks at your work without knowing what you meant it to mean.

[Ian kiprono @Carson_001]

Something I genuinely didn't anticipate: the cost of updating agents when the model underneath them changes.

The brief that worked perfectly on the previous version produces different output on the new one.

Sometimes better. Sometimes worse. Always different.

Model updates used to feel like good news.

Now they feel like a dependency upgrade.

You have to test everything.

[Ian kiprono @Carson_001]

It bothers me when people assume building in public is about visibility.

I find that backwards.

The visibility is a byproduct.

What building in public actually forces is that you can't be vague about what you're building.

Every week I post about Narrareach I find out if I actually understand what I'm building.

Often I find out I don't, as clearly as I thought.

[Ian kiprono @Carson_001]

Spent three hours debugging why my agent kept apologizing.

Turns out I'd trained it to say sorry for everything.

"Sorry, I can't find that order."

"Sorry, that feature isn't available."

"Sorry, I need to transfer you."

Customers started asking if something was wrong.

[Ian kiprono @Carson_001]

I find it interesting that the agents I've retired taught me more than the ones still running.

Each retirement was a decision.

This workflow no longer exists. This task is better done manually. This output quality isn't worth the cost.

Each one forced me to be honest about whether what I'd built was actually useful or just impressive.

The retired ones were honest feedback.

[Ian kiprono @Carson_001]

I find it telling that the hardest part of explaining my agent stack to someone non-technical isn't the technical part.

It's explaining why I trust it enough to let it run unsupervised.

And then explaining that I actually don't.

I check it every day.

The gap between automated and unmonitored is the one most people miss.

[Ian kiprono @Carson_001]

I find it strange that the pricing decision I agonized over most turned out to matter least.

Spent two weeks on tier structure.

What actually moved conversion was a single line of copy I changed in twenty minutes.

The work I spent the most time on and the work that had the most impact have been almost entirely different lists.

I'm still adjusting my judgment to account for that.

[Ian kiprono @Carson_001]

Something I find deeply underrated about building with agents:

The moment one of them produces something better than what you would have done yourself.

It happened for the first time about four months in.

The honest reaction wasn't pride.

It was a specific kind of discomfort I'm still trying to name.