What I built, what it was for, and what each one actually proves

Photograph a physical item, get back the verified catalogue record.

• Text recognition → candidate retrieval → ranking → eligibility rules → structured response
• 190 items catalogued through it
• Failed 6 of 7 real-world photographs on first contact; 5 root causes found and fixed
• 30-item reference set, measured at top-1 and top-5 — 30/30 after fixes, zero regressions
• Limitation, stated plainly: that set stayed green while the seven real photographs failed. It is a regression net, not a release evaluation
• 50 automated tests, duplicate protection, rate limiting, security hardening
• One case left deliberately failing: the record does not exist in the source catalogue

Four of the five fixes were not AI fixes. Most "the model is bad" turns out to be input handling, retrieval and business logic.

Designed, built, evaluated, operated

Decision-support software for choosing which AI use cases an organisation should fund and ship.

• Status: personal prototype. Built solo to model enterprise decision processes; never deployed inside an enterprise programme
• L0–L5 maturity model plus an 11-dimension AI operating profile layered on top
• 11-gate use-case lifecycle, from intake to production approval
• Use-case dossier carrying residual risk as an indication, never a verdict
• Evaluation and pilot-evidence layer feeding the gates
• Portfolio prioritisation with visible, movable weights — never a black-box score
• Vendor assessment as its own entity, deliberately unscored (a single number would fake precision)
• Adoption and realised-value tracking: plan versus actual variance, not judgement
• Around ten design decisions documented with the alternative rejected

The design principle throughout: the system never decides. It makes visible what you are risking.

Designed and built solo

An interactive briefing surface that makes an AI mandate legible to a leadership audience.

• Status: self-built briefing artefact using public and illustrative data. It has not been used in a client governance engagement
• Mandate gauge, value KPIs and a maturity view on one screen
• Filterable, sortable, expandable book of work
• An explicit RPA-versus-AI frame: when each is the right instrument
• Governance gates and a first-90-days plan
• Public and illustrative data marked distinctly, so nothing is passed off as verified
• Self-contained, works offline, light and dark

Half of consulting is translation. This shows how I structure an executive AI briefing — what leadership is asked to decide, and on what evidence.

Designed and built

The infrastructure I run my own work on: memory, tickets, sync, budget rules, generated views.

• A written memory policy: what is stored, where, and what is never stored
• Canonical source of truth separated from a generated human-readable view
• Idempotent generators, so three machines cannot fight over the output
• Silent git synchronisation across a Mac and two Windows machines
• Markdown ticket system with a CLI and a generated board — decisions survive the session
• Custom agent skills that enforce the working rules automatically
• Enforced cost discipline: exploratory fan-out requires stated scale and approval

At personal scale this exercises controls that also matter in enterprise AI operations — provenance, a single source of truth, auditability, reproducibility, cost policy. It does not demonstrate enterprise security, availability, access control or organisational adoption.

Designed and built

A test bench that scores competing prompt versions instead of arguing about them.

• Three summarisation prompt variants compared on the same inputs
• Assertion-based checks plus LLM-rubric scoring, where a second model grades the output
• Non-English test corpus, because quality claims rarely survive translation
• Provider configuration so the same suite runs across models

Prompt quality is testable. This harness turns prompt comparison into a repeatable, versioned test instead of an ad hoc chat session.

Designed and built

A phone-first tool that sorts incoming second-hand stock into categories from photographs.

• Status: working demo, live behind basic authentication and given to the customer at no charge — not a paid production deployment
• A three-category contract agreed with the customer before any model work
• Mobile-first web app; the vision-model call runs through a serverless proxy so the key never reaches the browser
• Roughly 2.5–3 HUF per image at the current model, on a deliberately capped key
• The decisive question was commercial, not technical: metered API plus engineering, or a flat subscription with a person in the loop
• At their volume the subscription came out ahead — cheaper, simpler, and it degrades gracefully when the shop is busy
• Awaiting the customer decision. Dedicated API billing, authentication and accuracy tuning are scoped but not built

An advisor who cannot recommend the boring answer is not much of an advisor.

Designed and built solo; delivered as a working demo for customer evaluation

A planning app for people with ADHD — currently in testing, and deliberately without an AI feature.

• Native SwiftUI, local-first, no account and no backend
• Third full design revision, driven by real use rather than opinion
• Calendar synchronisation, notifications, recurring schedules
• An obvious AI feature — automatic task breakdown — sits on an explicit "not doing this" list
• The reason is recorded: generating more items of variable quality is the worst thing you can do to an overwhelmed user

Knowing where AI does not belong in your own product is the same skill as knowing where it does not belong in someone’s business process.

Specification, architecture, review — AI writes the code