CONTEXT

PrepMyCV is an AI career-preparation web app for active job seekers in the EU and UK who send out ten or more applications a week. You add a job posting, get a free evaluation of how well you fit and whether the job is even worth pursuing, then pay per action for an honestly tailored CV and a grounded cover letter. I took it from first commit to production in under five months, covering product, design, and engineering direction.

WHAT I DID

I designed and built the whole platform, acting as product owner, architect, and operator, directing AI coding agents (Claude Code) through a spec-driven workflow: product briefs and PRDs first, then design specs, implementation plans, and adversarial audits (121 specs and 164 plans committed along the way).

• Product and UX: I created the Analog Command design system (industrial brutalism, tonal layering instead of borders) and a fully async experience where every AI action returns instantly and progress arrives through live task chips. Honesty is a design principle here: intake questions show the cost of skipping them, and weak matches get a clear view of what blocks them instead of false hope.

• Two-axis job evaluation: every job gets a free CV-fit and job-worth score, blending LLM judgment with deterministic skill matching (a self-hosted ML service with fine-tuned NER models over the 17,597-concept ESCO catalog). Calibrated over six weeks and twenty prompt versions against real data.

• Anti-fabrication tailoring: the pipeline analyses the CV against the job, asks up to 12 targeted intake questions, then tailors with deterministic gates that drop any claim it cannot trace back to the user's real history.

• Privacy as architecture: mandatory zero-knowledge encryption (AES-256-GCM under a 12-word phrase only the user holds), PII stripped before every AI call, GDPR export and erasure, EU hosting, and an EU AI Act assessment with visible AI marking on every export.

• The rest of a real product: credit-based payments with Stripe, a Chrome extension, deliverables in 25 languages, an admin console with full cost telemetry, and a guarded deploy process.

IMPACT

• Live in production, first commit to launch in under 5 months

• Around 180,000 lines of production TypeScript plus a Python ML service, built by directing AI agents rather than hand-coding

• 10,702 automated tests across 958 files, all passing, with zero type, lint, and import-cycle errors as standing policy

• Zero-knowledge encryption on every piece of career data, so even a database breach exposes none of it

• A documented, repeatable AI-native build method: specs, plans, audits, and a guard test for every incident lesson

Happy to demo it live and dig into any layer of it, in person or on a video call.

MY ROLE

Lead Product Designer

Product owner, architect, and operator

DURATION

February 2026 – present (live since July 2026)

SCALE

~180K lines of production code · 10,702 tests · 25 output languages

TOOLS USED

Claude Code, Stitch, Next.js, TypeScript, PostgreSQL, Python, Claude, GPT and Mistral models, Stripe, Playwright, Dokploy