$opinion

Opinion

What we have learned building this work, written down while it is still inconvenient. No thought leadership, no predictions — only things we have actually seen go wrong or right.

Everything else

index09 pieces
The eval set is the product
01AI

The eval set is the product

Teams keep shipping AI features with no way to tell whether a change made them better. Here is the harness we build first, every time.

Kabir Menon / Mar 2026 / 6 min
02Commerce

Replatforming is a revenue event

Treat a migration like a launch and you will lose orders. Treat it like a finance project and you might not.

Priya Deshmukh / Feb 2026 / 8 min
Your site is slow because of the third par
03Performance

Your site is slow because of the third parties

We audited twelve storefronts. In nine of them, the biggest single cost was a tag nobody could name an owner for.

Zoya Ahmed / Feb 2026 / 5 min
Most agents should be scripts
04AI

Most agents should be scripts

A taxonomy of tasks that genuinely need an agent, and the much longer list that does not.

Vikram Sethi / Jan 2026 / 7 min
What fourteen years of failed products tau
05Studio

What fourteen years of failed products taught us

Two shutdowns, one pandemic and a funding round that vanished. What we would do differently, specifically.

Arjun Raghavan / Jan 2026 / 9 min
Row-level security is where the incidents
06Data

Row-level security is where the incidents are

Every Supabase and Postgres incident we have been called into was a policy mistake, not a database one.

Farah Qureshi / Dec 2025 / 6 min
Headless is a content decision, not a spee
07Commerce

Headless is a content decision, not a speed one

If your reason for going headless is performance, you probably have a cheaper problem to fix first.

Kabir Menon / Dec 2025 / 6 min
Retrieval before prompting
08AI

Retrieval before prompting

Prompt engineering gets the attention. Retrieval decides whether the answer could ever have been right.

Nithya Balakrishnan / Nov 2025 / 7 min
09Studio

We stopped writing proposals nobody reads

What replaced our forty-page deck, and what it cost us to find out.

Rohan Dalvi / Nov 2025 / 4 min

One email, when there is something worth sending

No cadence, no calendar. A handful of times a year, and only when a piece is worth your time.

Unsubscribe in one click. We never share the list, and there are no tracking pixels in the mail.