Match Kit.

Sending CVs into the void

-- THE PROBLEM

In 2026, the frontend job market is saturated. Recruiters receive hundreds of applications per role, and most CVs are filtered by ATS systems before a human ever reads them.

The core problem is that most candidates have no idea how well their CV actually matches a specific job description. They apply broadly, hear nothing back, and have no signal for what to improve.

I wanted to build something that gives immediate, actionable feedback — a match score, identified gaps, missing keywords, and concrete rewrite suggestions — so candidates can stop guessing and start improving.

Designing for clarity under pressure

-- DESIGN CONCEPT

The design needed to feel sharp, trustworthy and fast — the opposite of a generic SaaS dashboard. I chose a dark-mode aesthetic with a single high-contrast accent colour, tight typographic hierarchy using DM Sans and DM Mono, and a deliberately minimal layout that puts the analysis results at the centre of attention.

Every design decision was driven by the emotional state of the user: someone anxious about their job search who needs clear, honest information — not noise. The result is a UI that feels clinical in the best sense — precise, readable and calm.

Design concept — dark mode, DM Sans + DM Mono type pairing

AI integration with a secure, server-side API proxy

-- ARCHITECTURE

The core technical challenge was integrating a third-party AI API cleanly and securely. Rather than exposing an API key client-side — a common mistake in AI demos — I built a server-side route handler in Next.js that proxies all requests to the Gemini API. The key never leaves the server.

The AI response is structured via a carefully engineered system prompt that forces the model to return strict JSON every time — match score, summary, strengths, gaps, keywords and before/after rewrite pairs. This makes the response reliably parseable and directly renderable in the UI without post-processing guesswork.

The result is a clean separation between the AI layer and the presentation layer — a pattern that scales to any AI-powered product.

The analyser in action — desktop & mobile

-- THE LIVE APP

The finished application — from the landing page through to a full analysis result with score, gaps, keywords and rewrite suggestions. Fully responsive across all screen sizes.

The live app — match score, gap analysis, keyword chips and rewrite suggestions.

The AI is only as good as the instructions you give it

-- PROMPT ENGINEERING

One of the most underestimated parts of building AI-powered products is prompt design. A poorly structured prompt returns inconsistent, unparseable output. A well-engineered one behaves like a reliable API.

For MatchKit, the system prompt instructs the model to act as a specialist CV analyst for frontend and tech roles, return only valid JSON with no preamble or markdown, produce rewrites that quote actual phrases from the CV, and ensure that strengths and gaps are specific to the job description — not generic advice.

This level of prompt precision is what separates a working prototype from a production-ready tool.

Prompt engineering · Structured JSON output enforced via system instruction

A production-ready AI product built in a single sprint

-- THE OUTCOME

MatchKit went from idea to deployed application in under a day. It demonstrates that AI integration does not require a complex backend or a large team — it requires clear product thinking, clean architecture, and disciplined prompt engineering.

This project reflects how I approach product work: identify a real problem, design a focused solution, and ship something that actually works.

Production deployment · Vercel serverless function with secure environment configuration