Building simon.kodu.se

Why

A portfolio should demonstrate, not just describe. Every architectural choice here is a talking point.

Architecture

• pnpm workspace monorepo — apps/web + packages/ui, packages/content, packages/config.
• Next 16 with cacheComponents — static shell streams instantly; work grid streams in after Velite-compiled MDX resolves.
• Velite MDX pipeline — work entries compiled to RSC-rendered components; rehype-pretty-code, remark-gfm, slug + autolink headings.
• Canvas particle system — 2D comet trails weave between DOM waypoints using catmull-rom interpolation, with reduced-motion, low-memory, and small-viewport opt-outs.
• Turbopack dev — sub-second HMR across the monorepo.
• Strict CSP — script-src 'self' 'unsafe-inline' enforced via middleware; MDX evaluated server-side so no browser eval.

AI fit-score

Recruiter pastes job description (or URL) → backend returns match_percentage, ranked strengths, gaps, and a 3-sentence summary.

• Route handler at /api/fit-score — gpt-5.5 via Chat Completions, response_format: json_object, max_completion_tokens: 1500. Param shape branches on model family (gpt-5 uses max_completion_tokens + default temperature; legacy gpt-4o accepted max_tokens + temperature).
• Prompt = candidate profile + JD + rules. Single source of truth (lib/fit-score.ts); swap providers = swap caller, not prompt. Stable prefix kept first for prompt-cache hits.
• Weighted scoring rubric — university courses count as real evidence, high-school courses as background only, courses with shipped projects weighted higher. Education-match bonus when ad asks for the candidate's programme; seniority cap only when ad explicitly demands years/full-time/senior.
• Baseline-familiarity rule — ubiquitous developer tools (GitHub, Excel, Slack, common IDEs) are assumed for any CS-bachelor candidate; never gap them as standalone. Git ⇒ GitHub, Jenkins ⇒ CI/CD inferred.
• Security-clearance handling — säkerhetsprövning / registerkontroll / försvarssekretess are employer-run vetting, not candidate skills. Swedish citizen ⇒ eligible; never gapped, never penalized.
• Newgrad-friendly detection + floor — regex flags ads with "nyexaminerad", "newly graduated", "från nyexaminerad till erfaren", etc. Post-LLM sanitizer applies a score floor (90/87/84/80 at 0/1/2/3 kept gaps) and a gap-removal bonus when the sanitizer drops false gaps.
• Explicit-de-prioritization rule — when an ad states attitude/learning matters more than exact tools, the merit list becomes flavor only — no gaps from it, floor lifted to ≥88.
• URL ingestion with SSRF guard — strips <script>/<style>/HTML, blocks private/loopback hosts (10., 127., 192.168., ::1, fc/fd…), 8s timeout, 500 KB read cap, 14000-char prompt cap.
• Strict JSON contract — { match_percentage, strengths[], gaps[], summary } (strengths 2–4, gaps 0–3). Parse failure → 500, no leaked prose. Preview/debug mode attaches a full scoring trace (llm_raw, jd_flags, gap_decisions, score pipeline) so every adjustment is auditable in the UI.

Performance budget

Lighthouse CI gates every PR: perf ≥ 0.9, a11y ≥ 0.95, best-practices ≥ 0.95, SEO ≥ 0.95, LCP ≤ 3.0s, CLS ≤ 0.05.

Tooling

• Biome — single-binary lint + format.
• Vitest + @testing-library/react — component unit tests.
• Playwright + @axe-core/playwright — e2e and a11y sweeps.
• Husky + lint-staged + commitlint — conventional commits enforced pre-commit.