— THINKING

Notes on products,
AI and decisions.

Working points of view shaped by building, evaluating and changing my mind.

SIGNAL

PROBLEM

DECISION

PRODUCT

LEARNING

01 / HOW I THINK

Three principles I keep returning to.

They are not universal rules. They are working defaults that help me move from ambiguity to a deliberate product decision.

Start with the decision

Understand what someone needs to decide before designing the interface.

Test the riskiest belief

Use the prototype to reduce the uncertainty that could invalidate the idea.

Reveal complexity progressively

Offer a useful first layer, then reveal depth when the user asks for it.

02 / DECISION NOTES

Ideas earned through the work.

Each note connects a product belief to evidence from something I built.

TESTING IDEAS4 MIN READ

A prototype should test the biggest question—not just show the easiest feature.

A good-looking demo proves that something can be built. A useful prototype helps us decide whether it should be built.

It is tempting to begin with the most visible screen because it quickly makes an idea feel real. But the screen may not be the biggest unknown. The real question could be whether the information is reliable, whether the advice is genuinely helpful, or whether people understand what to do next.

Before building, I now ask: what would make this idea fail? PRISM had to show that scattered research could become one clear meeting brief. Smart Finance had to show that guidance could be more useful than simply displaying expenses.

WORKING PRINCIPLETest the question that could change whether the idea is worth building.

Evidence from PRISM

03 / PATTERNS ACROSS PROJECTS

Different domains. Recurring decisions.

The projects changed, but several reasoning patterns appeared repeatedly.

PROJECTPRINCIPLEEVIDENCE IN THE WORK

01 · PRISMSeparate before combining

Specialist research agents and structured outputs

02 · SMART FINANCEPreserve user control

Manual categorization and feedback around AI guidance

03 · APEX F1Reveal depth progressively

Season scan → archive → race replay

04 · SPOTIFY ANALYSISKeep metric grain explicit

Raw source totals separated from the cleaned model

05 · SENTIMENT ANALYSISEvaluate behavior, not headlines

Macro F1, class recall and separate irony detection

04 / STILL EXPLORING

Questions I have not finished answering.

These are the tensions I want to keep testing through real products and teams.

Where should AI recommend, and where should it only explain?

How much evidence does a user need before trusting an output?

How should product teams communicate model uncertainty without creating paralysis?

When does personalization create value—and when does it add unnecessary complexity?

— A WORKING POINT OF VIEW

Built through practice.
Open to revision.

These ideas are shaped by what I have built, what the evidence showed and the moments when the original assumption did not survive.

— THE NEXT DECISION

Have a complex problem
worth simplifying?

I’m exploring entry-level Associate Product Manager and AI Product opportunities where I can combine user understanding, analytical thinking and technical execution.

Email me LinkedIn View Resume