Ankur Kulkarni

I run the same loop on every product: research, hypothesis, build, test, measure, decide, and repeat. From KnewsIQ onward, language models shorten each stage, and I added a stage for checking what they produce.

1. 01

   Research and hypothesis

   On KnewsIQ, a model produced the first analysis of my interview notes, competitors’ reviews and my notes on competitors’ products, which cut collation from days to minutes. At each change of target user I refined the target user, personas and jobs to be done with a model, and that work set the taxonomy.

2. 02

   Build and validate

   On KnewsIQ I tested each new direction as two or more AI-built mockups or landing pages, showed them to target users and iterated until the problem, user and solution were clear. Only then did I build the change. I build MVPs with AI coding agents: KnewsIQ on no-code tools (WeWeb and Xano) with AI assistance, and HoldOff entirely with Claude Code.

3. 03

   Test and measure

   I read usage data and feedback with a model’s help to form hypotheses, design tests and build them, and use the same loop to investigate why a metric is moving. On KnewsIQ this produced the test of four email send times, in which weekday lunch led on open and click rates.

4. 04

   Check every model output

   I check everything a model produces, from code to analysis, and most of all output that reaches users. On KnewsIQ that meant a weekly sample of Story Cards and rejected articles judged against a rubric, with failures grouped and a guardrail added for each group before the next round. On HoldOff, an independent tester role writes its own tests using only the spec.

5. 05

   What I keep for myself

   I interview users, use competitors’ products myself, check a model’s analysis against my own notes, and decide what to build. I check the analysis because models tend to merge separate points into one conclusion. On HoldOff the builder lists its own choices for my sign-off.

How I run the work

Acceptance criteria

Written on every documented product: RoughSheet, FindX (for every feature) and KnewsIQ.

Experiments

Lean Startup, Agile and Scrum across RoughSheet, the consulting engagements and KnewsIQ. On RoughSheet that meant 200+ experiments over five years.

Cadence

Three-week sprints on KnewsIQ, with the next sprint’s work chosen at the end of each. From late April or early May 2026, acquisition, conversion and retention work ran in parallel within each sprint.

Priorities

On KnewsIQ, content quality ranked above other considerations when choosing what to build next.

New technology

On KnewsIQ I assessed new technologies the way I assess a competitor’s change: whether it affects the product, what a similar build would gain, and whether that justifies the time and effort before it goes into an existing sprint. Two examples were agentic workflows against direct API calls, and fine-tuning against rubric-based evaluation of selected records and logs run with LLMs.

Decision rules

On HoldOff, conflicts between sources follow a fixed rule: the spec decides behaviour and wording, the mockups decide layout, and the design system decides visual tokens.

Code

I am a non-expert coder with working knowledge of JavaScript, Python, SQL and the terminal. I ship by directing AI coding tools and checking what they produce.