You can just build things.

Ecommerce catalogs are full of ambiguous visual attributes that are easy for humans to see and painful for systems to classify consistently.

The useful unit is not a magical label. It is a visible taxonomy decision with enough reasoning that a merchandiser can challenge it.

A classification workflow that combines visual analysis, structured attributes, scoring, and recommendation logic without hiding the taxonomy.

A product direction where AI output becomes inspectable business infrastructure instead of a black-box content generator.

Trust grows when people can see the system's categories, not just the system's confidence.

Business geography often lives in spreadsheets, screenshots, or ugly default maps that fail to tell the strategic story.

A map becomes useful when it is designed as communication, not when it merely proves coordinates exist.

Address ingestion, geocoding, styling, export logic, and designed map outputs built for founders who need to explain distribution, growth, or coverage.

A practical product wedge around turning operational data into something people can understand in one glance.

Design is not decoration here. It is the difference between data storage and decision support.

Most baby tracking tools treat exhausted parents like data-entry clerks with unlimited attention.

The best interface for tired people is the one that assumes partial focus, repetition, and tiny windows of time.

A lightweight product exploration around quick capture, practical summaries, and caregiver handoff.

A proving ground for building personal software that respects the user's actual energy level.

Real constraints make product taste sharper. Sleep deprivation is a brutal UX reviewer.

Small businesses lose hours in follow-up, quoting, scheduling, and checking whether event details are actually complete.

AI becomes valuable when it attaches to the operational moments where work already gets dropped.

CRM-aware workflows for opportunity review, quote and invoice drafting, French customer communication, and logistics completeness checks.

Proof that the same AI principles apply outside software labs, where the work is concrete and the mistakes are visible.

Automation earns trust when it helps the owner move faster without pretending to own the customer relationship.

Agent demos often look smart because the task is staged. Production work has messy tools, partial context, and consequences.

The core question is not whether an agent can act. It is when it should act, how it knows enough, and where it stops.

Experiments with assistants, tool calls, execution loops, context loading, and human review points.

Sharper intuition for where agents help, where they create risk, and how to describe them without mythology.

Agents are less like employees and more like workflow muscles: powerful when attached to a skeleton.

A serious company cannot adopt GenAI through vibes. It needs shared understanding, clear business direction, and practical enablement.

The adoption problem is partly technical, but mostly organizational: people need language, judgment, and confidence.

Public-safe summary: GenAI Principal at N-SIDE, helping shape business vision and upskill around 200 employees.

A practical adoption role focused on making AI understandable, useful, and connected to business value.

AI adoption works best when employees understand both the opportunity and the limits.