How this site is built
The portfolio is itself a shipped product, so it gets the same treatment as the case studies: what it runs on, the decisions behind it, and the numbers it scores.
Measured
Google Lighthouse against harveys.vercel.app, 2026-09-17. Performance, accessibility, best practices, SEO. Desktop is one run; mobile is the median of five, because simulated phone runs vary by several points.
desktop
- 100
- Performance
- 100
- Accessibility
- 100
- Best practices
- 100
- SEO
LCP 0.6 s · CLS 0.001 · TBT 0 ms
mobile
- 90
- Performance
- 100
- Accessibility
- 100
- Best practices
- 100
- SEO
LCP 2.1 s · CLS 0 · TBT 300 ms
Desktop scores 100 on all four. The one number below 95 is mobile performance, a median of 90 across five runs. Largest contentful paint is the portrait at 2.1 seconds, inside the 2.5 second target, after it lost its entrance animation: an opacity fade on the largest paint made the browser wait for hydration before counting it. What holds the score now is about 300 milliseconds of main-thread time while React hydrates the page on a throttled phone. That is the cost of the interactive parts, and it is listed here rather than hidden.
The Ask box, graded
A fixed set of 33 questions, 28 the site can answer and 5 it must decline. Retrieval is graded on whether the right section reaches the top 5 of 107 chunks. An answer is graded on whether it cited that section, or refused when it should have. Both modes are scored, because the box runs with or without a model.
Retrieval
- 100%
- Hit at 5
- 78%
- Mean reciprocal rank
BM25, deterministic, no network
Answers, no model
- 68%
- Grounded in the right section
- 100%
- Refused when it should
19/28 grounded, sentences selected from the page
Answers, with a model
Last run hit the provider's free daily limit partway through, so most answers came from the fallback. That is not a measurement of the model, so no number is shown.
The gap between the middle column and the right one is the whole argument for using a model here. Selecting sentences is honest and free but literal: ask what Harvey built at The Asset and a lexical matcher cannot tell the company from “S3 for assets”. The model closes that gap, and on a free tier it is also the part most likely to be unavailable, which is why it is the part allowed to fail. Measured 2026-09-17.
Stack
Next.js 16
App Router, static generation for every page, image optimisation, one deploy per push.
React 19 with the React Compiler
No hand-written memoisation. The compiler does it.
Tailwind CSS 4
Design tokens live in one CSS file and every component reads them.
No animation library
Every entrance, the skill-map path draw and the section reveals are CSS keyframes; the count-up is twenty lines of requestAnimationFrame. Removing the library took 48 KB of JavaScript off every page.
Native CSS scroll-driven animation
Section reveals use animation-timeline: view(), so they can never strand content invisible and cost no JavaScript.
Phosphor Icons
One icon family, one stroke weight, no hand-drawn SVG paths.
Vercel
Linked to the GitHub repository. Pushing to main is the release.
No fixed AI vendor
The Ask box takes whichever provider key is set, and answers by selecting sentences from the page when none is. The retrieval and the citations are the same either way.
Decisions worth explaining
01
Two content files, no CMS
Everything on the site is in content/site.ts and content/projects.ts. A case study is an object. Adding one is a commit, and the same file drives the skill map, the sitemap and the Open Graph tags, so they cannot disagree.
02
Every skill is tied to a project that used it
The skill map is not a list of things I have heard of. It is computed from the stacks of the featured case studies, so a skill only appears if a project on this site actually used it, and hovering it shows you which one.
03
Lines behind words, always
The skill map draws a curve per project-to-skill link. Curves live on a lower layer and every label paints an opaque background, so a line terminates at the edge of a word rather than cutting through it. Dimming is done by ink colour, never element opacity, because opacity would make the label backgrounds translucent.
04
Screenshots, not mockups
Every project image is the real thing, captured from the live site or the store listing. Where a product is on the App Store, the listing itself is shown. Where a live app allows it, the case study embeds it.
05
Motion that says something, and never gates the paint
The hero grid pulses, the proof numbers count up, active skill links carry a travelling pulse. Each has a job: attention, arrival, direction. All of it collapses to static under prefers-reduced-motion. The portrait, the page's largest paint, has no entrance at all: an opacity animation on it made the browser wait for hydration before counting it painted, which on a throttled phone was most of the largest-contentful-paint.
06
In progress means in progress
Unshipped work is labelled as such and dated. Nothing on the site claims a result that has not happened.
07
The agent retrieves, cites and refuses
The Ask box is not a chat window over a language model. BM25 ranks the site's own content, the top five chunks go to the model as numbered sources, and every claim carries a citation linking to the section it came from. A question the site cannot answer gets one fixed sentence. The whole thing is graded against a 30-question set, and the scores above are the ones it got.
08
The model is the part that is allowed to be missing
Retrieval, citation and the decision to refuse are mine and cost nothing to run. Generation is a swappable provider, and with no key set at all the same sources are answered by selecting the sentences that match the question, unedited. So the box cannot hallucinate when it is running that way, cannot go dark when a free tier runs out, and cannot hold the site hostage to one vendor's pricing. A model makes an answer read better. It is not what makes it true, and the eval above scores both modes so the difference is a number rather than a claim.
09
The old site is kept, not hidden
The 2020 template version is archived under /legacy, banner on top, excluded from search. Anyone can compare.
Deliberately absent
- No analytics or tracking scripts. Nobody is measured for reading a CV. Questions typed into the Ask box go to whichever model provider is configured, are not logged here, and are answered without one at all if none is set.
- No third-party fonts loaded at runtime. Geist is self-hosted through next/font.
- No third-party icon CDN. The stack marks are local SVGs.
- No 3D, no particle background, no custom cursor. The work is the effect.
- No cookie banner, because there is nothing to consent to.
Source lives in a private repository. The archived 2020 version is at /legacy. The workflow used to build it is public: claude-setup, and this site is one of its case studies.