A fitness tracking website, a member application and a separate staff console, built from a supplied Illustrator identity. This page walks through the decisions and shows the result, with the measurements that back each claim.
Trackly records the training, food, water and measurements a person enters, and shows what changed over twelve weeks. There are three surfaces, and they are deliberately different from one another.
A public website
The problem
Somebody deciding whether to bother needs to know what the thing actually records and what it refuses to do.
What we did
Marketing pages that each carry a 'what this does not do' panel, written in the same plain language as the product. No testimonials, no user counts, nothing that cannot be shown.
A member application
The problem
Logging has to be fast enough to do between sets, and the figures have to be trustworthy enough to act on.
What we did
A dense dashboard where every number is recomputed from the member's own rows on each page load. Nothing is stored as a dashboard total, so editing a session changes every figure that reads it.
A staff console
The problem
Reviewing a member's submitted file needs different tools and different access from using the product.
What we did
Its own address, its own session with a shorter life, its own sign in, and a permission check by name on every page and every action. The public domain answers 404 for it.
Every member account and every record in the live build is fictional demonstration data, generated by a seed script so the same figures appear on every screenshot. The product says so on the dashboard, in the footer and in the admin console. No real person is described anywhere in it.
02
The identity, taken rather than guessed
The client supplied one Illustrator file. The first job was to read it properly instead of picking colours off a screenshot by eye.
The source artwork was opened in Illustrator through a script, not by hand. That script read the fonts, the swatches and the fill of every path, then exported the mark and the wordmark as vectors. A second pass built a pixel histogram of the exported artboard, so two independent methods had to agree before a colour was written down.
The blue turned out to be pure #0000FF, which is unusual and would have been guessed wrong from a screenshot. The wordmark is outlined vector art in the source, not live type, so it is reproduced here as paths and is never re-set in a font.
The one problem the extraction found
The headings in the source are set in TT Fors Trial. A trial licence cannot ship on a public website. Rather than substitute something quietly, the product uses Archivo, an Open Font Licence face with the same tall x height and flat terminals. Archivo is a variable font with a width axis, so the condensed cut that carries every figure and the normal cut that carries reading text are one family rather than a pairing. The wordmark is unaffected because it is artwork, not type.
The mark holds down to 18 px, which is the size it runs at in the application rail.
#0000FF
Trackly blue
The field, the figure that is going the right way, every active state. White on it is 8.59 to 1.
#0B0B0C
Ink
Type, every hairline, and the dark field. 19.3 to 1 on white.
#FFFFFF
Paper
The ground. There is no second surface colour, because nothing in the product is a card.
#D81E35
Red
A figure going the wrong way, and nothing else. It never appears as decoration.
#DEDEE1
Rule
Every hairline in the product. It is the only thing that separates one block from another.
#5E5E63
Secondary text
Reading text under a heading. 5.9 to 1 on paper, so it carries body copy at any size.
#F2F2F4
Marker
Not a surface. It marks a value or a state, a chart ground or a day already past, and it is the faintest step the eye reliably finds.
03
An icon set with one hand
64 icons, all on a 24 unit grid with a 2 unit stroke, square caps and mitred joins. Those terminals are not a style choice: they are how the Trackly mark itself ends.
alert
approve
arrow
audit
bell
bike
calendar
chart
check
chevron
clipboard
clock
close
cycling
download
droplets
dumbbell
edit
export
file
filter
flame
food
heartpulse
home
interval
link
lock
logout
macro
medal
menu
message
minus
mobility
plus
privacy
protein
queue
refresh
reject
repeat
rest
revise
role
rowing
run
running
scale
search
settings
shield
stopwatch
streak
strength
swimming
target
trash
trend
trophy
upload
user
users
walking
The client's starters were kept
The problem
The brief supplied 44 editable starter icons. Throwing them away would have lost work that was already paid for.
What we did
All 44 were carried through and re-gridded. The stroke weight, the caps and the joins are applied by one component at render time, so the whole set changed hands in one edit rather than 44.
Four were redrawn
The problem
A target drawn as three concentric rings, and a medal drawn as a rounded pill, read as stock icons and fight the brand's flat geometry.
What we did
They were redrawn against the mark's own construction. The target became a bar reaching a marked line, which is also what the product actually shows on a goal.
Twenty were added
The problem
The starters had no pictograms for the training types, the macro groups or the review states the product needed.
What we did
Twenty more were drawn on the same grid, so a session type, a macro and a review decision all read as the same family.
The Trackly mark is not in the icon set. Redrawing it on a 24 unit grid would change it, so it lives at its own size as artwork and is used through its own component. Nothing in the product traces, approximates or re-draws it.
04
What the references actually decided
A pool of real work was gathered and judged before any screen was drawn, because a direction invented from the brief alone lands on the average of everything, and the average is what reads as machine made.
Searches8Chosen as subjects, not as formats.
References594All of them downloaded and catalogued.
Contact sheets24Every one of them looked at.
Directions kept1The record book. The rest were refused.
The searches were run against the subject rather than the format. Searching "fitness app ui kit" returns template kits, and a direction built on template kits is the thing clients recognise as generated. Searching the subject, which here is sport, records and measurement, returns work with a point of view that nobody else in the sector can borrow.
The negative register, and why it came first
One search was run deliberately to find the trap. Fifty three fitness app kits came back and they are interchangeable: rounded cards on a pale ground, a lime or orange accent, progress rings, an icon in a tinted rounded square above a bold title and a grey subtitle, soft shadows under everything. The first version of this interface was a well built example of exactly that, which is why the direction below is defined first by refusal.
477A football stat sheetTaken: a tiny label in capitals directly above a very large figure, in a row of four. It is the opening of every page in this product.
489A record book spreadTaken: a ranked table set as proper typography is enough to be a page on its own. It does not need a card around it.
490A year in numbersTaken: order a set of figures by magnitude and let that ordering be the layout.
485A list of broken recordsTaken: the ledger row. Name on the left, figure on the right, one hairline between them and nothing else.
378A contents pageTaken: an index. Small figures against large words, so a list of five things is the page rather than a section inside it.
158A broadcast half time tableTaken: the two sided row, figure then label then figure, which is how any comparison is shown here.
Numbers are catalogue entries in the research folder, which records the source of every reference. What a reference contributes is a mechanic, a register or a piece of typographic courage. No layout, asset or composition was reproduced, and the product carries no Tailwind, no component library and no borrowed code.
05
The design system
Tokens first, then components, then pages. The rule the whole system follows is that a training log is a record book, so the product is built from flat fields, large figures and ruled rows, and from nothing else.
Type
Display
Archivo 700, tracking minus 0.035 em, uppercase, line height 0.96
1,998
The figure block. Archivo 700, tabular numerals, the loudest thing on any screen
Body copy runs at 1.125 rem with a 1.55 line height and a 58 character measure.
Archivo 400
kcal · km · min · sessions
Archivo 700 at the normal width, in capitals and letterspaced. Every figure carries one
The rules that do the work
Depth comes from a three step stack, canvas to surface to panel, not from shadow. There is one shadow in the whole product and it is on a floating menu.
Hierarchy is size and weight. Opacity is never used to make something quieter, because a softened tone is a contrast cut wearing a disguise.
One repeating device: a three unit blue bar, taken from the bar in the mark. It heads a section and marks the active item, and it is the only ornament that repeats.
One held back device: a 45 degree corner cut, taken from the mark's arms. It appears once, on the sign in screen, and it runs off the edge rather than floating inside it.
No gradient on an interface surface. The brand gradient is used on marketing panels and on the article covers, never behind data.
No stock photography, on purpose
The problem
The client's brand board leans on athletic photography, so the obvious move was to buy some. The obvious move is also how every fitness product on the internet ends up looking identical: a stranger mid lunge, shot in a gym nobody trains in.
What we did
The public pages are built from type, data and screenshots of the real product instead. A reader deciding whether to bother is better served by the actual workout list than by a model. If the client wants photography later it has a place to go, on the hero and the audience cards, and the layout already holds it.
No illustrative people anywhere
The problem
Demonstration data needs names and faces to feel real, and invented faces are either stock photographs of real people used out of context or generated ones, which is worse.
What we did
Twenty four fictional members with names and records, and no portraits at all. The member avatar is initials on one of three brand colours, derived from the account id, so it is consistent without depicting anybody.
Charts are hand drawn in SVG
There is no charting library. Three reasons: every axis can carry its own unit, the marks use the brand's flat terminals rather than a library's round ones, and the page ships no chart runtime. The line chart, the bar chart, the stacked split, the sparkline and the twelve week activity grid are all written against the same token set, and every one of them prints its unit in a caption underneath as well as on the axis.
Five chart types on one page, each with its unit stated on the axis and again in the caption.
06
Three compositions, one brand
The brief asked for three meaningfully different primary compositions. They differ in structure and density, not in colour palette, which is what keeps them recognisably the same product.
Public: editorial, wide measure, one idea per screen
Marketing
The home page end to end. The only backgrounds on it are the two full width bands; everything else is type on paper, separated by rules.
Member: a stat line, then ruled sections
Application
The trend chart takes eight of twelve columns. Below it every section runs the full width, because a half width block beside a taller one leaves a hole.
Admin: tables and queues, built for scanning
Staff console
The console leads with what needs a person today rather than with a vanity figure. There is no payment system, so no revenue is shown.
07
The member journey
Four steps from a stranger to a trend worth reading. Each screen below is the real one.
01 Register. The password rule and how it is stored are stated under the field.
02 Four questions. Every one can be skipped or changed later.
03 Log a session. Sets are recorded one by one, with warm up sets marked.
04 Read what changed. Every figure is added up from the rows at page load.
A trend, not a reading
The problem
A single morning weight moves with hydration, salt and the time of day, so a daily number makes people chase noise.
What we did
The weight chart draws a seven day centred average over the member's own entries, and the caption says exactly that. The raw entries stay in the table underneath.
Personal bests are computed, never stored
The problem
A stored best goes stale the moment somebody edits or deletes the session behind it.
What we did
Bests are derived from the logged working sets every time the page loads. Warm up sets are excluded, and the estimated one rep max is labelled as an estimate with the formula named.
A missed session is a record
The problem
A calendar that only shows what happened cannot tell a rest day from a day somebody skipped.
What we did
A planned session that did not happen is stored with that status and shown on the calendar as a dashed outline, so the member can see the difference.
08
Uploads and the review lifecycle
Ordinary logging never needs anybody's approval. The review workflow exists for files exported from another app, where a person checks the data before it is accepted.
What the member sees. The lifecycle bar is filled to where the file has actually reached.
What the reviewer sees, on the separate console.
Two reviewers cannot take the same file
The problem
A queue where two people can claim the same item produces two decisions and one annoyed member.
What we did
Claiming is a conditional update that only applies while the row is still unclaimed. The second reviewer is told it has already been taken, and nothing changed.
A decision needs a reason a member can read
The problem
A rejection with no reason produces a support ticket, and a reason written for the team is no better.
What we did
The form refuses a reason under ten characters, says the member will read it, and stores it with the file as well as sending it. Internal notes are a separate, clearly marked thing.
Progress photos stay shut
The problem
A reviewer checking a CSV has no business opening somebody's progress photograph.
What we did
Photos are only opened if the member switched that on, which is off by default. Where it is off, the console says so and the reviewer judges the submission on the file details.
09
The admin console
Its own address, its own sign in, its own session. It shares one database with the member application, which is why a decision made here shows on the member's own page immediately rather than after a sync.
The queue, filtered by state, with the parser result on every row.
Members and staff. Staff never see a password, because none is stored in readable form.
The audit log. Written at the moment of the action and not editable from the console.
Settings, including a permission table read from the same definition the code enforces.
10
Database and architecture
27 tables, 2 migration files, one Postgres. The schema is the product: if a figure cannot be derived from these rows, the product does not show it.
What is in the live database
Member accounts
25
All fictional, all seeded
Workouts
1,079
8,137 individual sets under them
Food entries
7,841
across 1,430 member days
Hydration logs
5,768
Measurements
512
Goals
63
Uploads with review history
42
Audit events
51
The reconciliation check
A claim that the dashboard agrees with the rows underneath it is worth nothing unless something checks. A script adds the raw rows up for one member and compares the result with what the dashboard query returns, for the window 2026-10-20 to 2026-11-18.
Figure
Dashboard says
The rows say
Result
Sessions in the window
15
15
Agrees
Minutes in the window
881
881
Agrees
The same script runs eight integrity checks that should all return zero: orphaned workouts, orphaned sets, sessions dated in the future, negative durations, negative energy, duplicate measurement days, accounts without a password hash, and any password not stored in Argon2 form. All eight returned zero on the last run.
11
Permissions
Five roles. Permissions are checked by name on the server for every page and every action, never by comparing a role string at the call site, so adding a role cannot quietly grant something nobody meant to grant.
Permission
support
reviewer
admin
owner
users.read
users.moderate
—
—
users.role
—
—
—
reviews.read
—
reviews.decide
—
content.read
content.publish
—
—
catalogue.write
—
—
support.read
—
support.reply
—
reports.read
audit.read
—
—
settings.write
—
—
—
This table is rendered from the same definition the application enforces, so it cannot drift away from what is actually true. A member role appears nowhere in it, and a unit test asserts that.
12
Security decisions
Each of these is a choice with a reason, not a checklist item.
Passwords
The problem
A stored password is a liability the moment a database leaks.
What we did
Argon2id hashes with a 19 MB memory cost. No password is readable by anybody, including staff. A staff member can send a reset link to the member's own address and nothing else.
Sessions
The problem
A single session that works everywhere makes the admin console only as safe as the weakest member device.
What we did
Two cookie names and two scopes. A member session cannot open the console and an admin session cannot open the member app. Admin sessions last eight hours, member sessions fourteen days, and both are httpOnly, secure and same site.
Ownership on every read
The problem
A detail page that fetches by id alone will happily show one person another person's records.
What we did
Every member query filters on the owner as well as the id, so asking for somebody else's address returns 404 rather than their data. An end to end test does exactly that and asserts the 404.
Uploaded files
The problem
A file at a public address is one guessed URL away from being somebody else's.
What we did
Bytes are stored in a database row, not in object storage behind a public link. There is no address to guess, and every read goes through a query that checks who is asking.
Sign in attempts
The problem
An unlimited login form is a password guessing service.
What we did
Failed attempts are counted per address and per hashed network address, and the eighth failure in fifteen minutes stops the form. The message never says which half was wrong.
The audit log
The problem
An action nobody recorded is an action nobody can answer for.
What we did
Every staff action and every state change a member cannot undo is written with the actor, the action, the record and the time. The network address is stored as a one way hash so events can be matched without identifying a location.
What has not been done: no penetration test has been carried out, and no independent security review. The checks above are the ones built into the product and covered by its own tests.
13
Email
Thirteen messages wired to something that sends them, one layout, and an honest state when no provider is configured.
What is built
One layout renders every message: tables and inline styles, the real mark as vector, the brand's own colours, and a health disclaimer in the footer. Thirteen messages are wired to something that sends them: email confirmation, password reset, password changed, upload received, approved, rejected and needs changes, the reviewer alert, the contact receipt and its staff alert, goal reached, account restricted and deletion requested. Two more pieces of copy are written and deliberately not wired. The welcome is delivered inside the application as a notification instead, and the weekly summary would need a scheduler this build does not have, so nothing sends it and nothing pretends to.
Two drivers are implemented. Resend sends for real. A file driver writes each message to disk, which is how the templates were checked without sending anything to anybody.
Every send is written to a log table with its outcome, so a delivery failure is visible in the console rather than silent.
What happens with no provider
The deployment currently runs with the mail driver switched off, because no Resend key was available to this build. That state is not hidden:
The reset form says reset by email is switched off, and sends nothing.
The settings page says the same and disables the resend button.
The contact form still records the ticket and says the receipt cannot be sent.
Each attempt is logged as skipped, with the reason.
Setting three environment variables turns all of it on. Nothing else changes, because the flows are already built and tested against the file driver.
14
Search and analytics, and what is honestly not on yet
The implementation is complete and the configuration is not, because the accounts were not reachable from this machine. Both halves are described here rather than one.
Built and working
A title and a description written per page, not generated from a pattern.
A canonical link on every public page.
An XML sitemap generated from the published articles and the real routes.
A robots file that keeps the member app, the console and the API out.
Open Graph and Twitter card metadata with an image.
Structured data on articles and on the questions page, describing what is actually there.
Permanent redirects from the older addresses in the brief to the ones in use.
One documented event layer with a fixed list of nine event names.
A consent panel that loads nothing until a visitor chooses, verified by a test that watches the network.
Configured but not active
Google Tag Manager is the only container the page would load, and GA4 is configured inside it rather than beside it, so there is no way to double count a page view. The container id comes from one environment variable. With no id set, nothing loads and the consent panel does not appear, because there is nothing to consent to.
The event helper accepts nine named events and a small set of non identifying parameters, and silently drops anything else. Nothing about a member's body, training or food can reach it, by construction rather than by policy.
No Search Console property has been verified, no GA4 property or stream exists, no Tag Manager container has been published, and no Google Ads conversion action has been created. Nothing has been indexed. These are stated as not done rather than described as pending.
15
What was measured
Numbers on this page come from files written by tools that looked at the running product. Where a check has not been run, it says so instead of guessing.
Unit tests
40 of 40
Formatting, units, streaks, permissions
End to end tests
50 of 51
0 failed, 1 skipped, run in a real browser
Screenshots captured
59
51 at 1440 px, 8 at 375 px, 0 unexpected statuses
Contrast failures
1
out of 5,555 text nodes on 40 routes
The contrast sweep
A contrast failure does not collide with anything and does not leave its box, so looking at a page will not find one. A script drives the real browser, walks the DOM of every route, takes the computed colour of each piece of text and the effective background behind it, composites any transparency, and scores it against the WCAG AA threshold for its own size and weight.
Where text sits on a gradient it is measured against every stop behind it and scored on the worst one, because an average colour exists nowhere on the screen.
The first run of this sweep found 45 failures that nobody had noticed, all from two causes: a muted grey chosen against white that fails on the grey canvas, and an outline button inheriting the ink colour on an ink card. Both were fixed at the token level. The current run measures 5,555 text nodes across 40 routes with 1 below the threshold.
What the end to end tests actually do
Register, land on setup, finish it, and reach a working dashboard.
Refuse a short password, a mismatch and an unticked terms box.
Prove the register and reset forms do not reveal whether an address exists.
Create, read, edit and delete a workout, and check the figures change with it.
Record a strength session and check the working volume is computed correctly.
Add food and assert the headline total equals the sum of the listed rows.
Press a hydration button and assert the figure moved by exactly that amount.
Save two measurements for one date and assert the second replaced the first.
Download the CSV export and check its header against the columns shown.
Have one member ask for another member's record and assert a 404.
Check a support role cannot open the audit log or the settings page.
Take a file through submitted, under review, needs changes, resubmitted and approved.
Publish an article from the console and fetch it from the public site.
Assert no analytics request is made before a visitor consents.
Check every public page for one h1, a canonical link and a description.
Crawl every internal link on the public pages and assert none is broken.
Check four pages at 375 px for sideways overflow.
Retries are switched off in the test configuration. A test that only passes on the second attempt has found something, and hiding that behind a retry is how a flaky product ships.
16
Deployment
One deployment, two hosts, one managed Postgres.
The versions actually installed
next
15.5.29
react
19.0.0
typescript
5.7.3
drizzle
0.45.4
drizzle Kit
0.31.11
postgres Driver
8.13.1
pglite
0.2.17
argon2
2.0.2
zod
3.24.1
resend
4.0.1
playwright
1.56.1
vitest
2.1.8
sharp
0.33.5
Read from package.json rather than typed out, so this table cannot go stale.
How it runs
Database. A managed Postgres on the free plan, TLS enforced, its own instance rather than one shared with another project. Migrations are applied by a script before the new build serves, never automatically at boot, because a web process racing itself to migrate during a rolling deploy is how a table goes missing.
Two hosts. Middleware routes by host. The admin host serves only the console. The public host answers 404 for every admin address, verified against the deployment.
Environment. The application refuses to start in production without a database URL, with a session secret under 32 characters, with a non https site URL, with the file mail driver, or with seeding enabled.
Health. A health endpoint reports which database engine is actually serving, whether mail is configured and whether analytics is configured, so none of those has to be taken on trust.
17
Evidence ledger
What is verified, what is built but cannot be verified from here, and what is blocked. Nothing is marked done because it was written.
State
What
Evidence
Verified
Public website, every page
All pages answer 200 with one h1, a canonical link and a description. Checked by test and captured as screenshots.
Verified
Register, sign in, sign out, reset request
Driven in a real browser, including the cases that must fail and the two forms that must not reveal whether an address exists.
Verified
Workout create, read, edit, delete
Driven end to end, with the derived figures checked after each change.
Verified
Nutrition, hydration and measurement totals
Tests add records and assert the headline figure equals the sum of the listed rows.
Verified
Upload lifecycle through to approval
A file taken through submitted, under review, needs changes, resubmitted and approved, across two browser contexts.
Verified
Cross account denial
One member asks for another member's record by its exact address and receives 404.
Verified
Role restrictions in the console
A support role is turned away from the audit log and the settings page, and a reviewer from the catalogue.
Verified
Admin console on its own host
The admin host serves the console and the public host answers 404 for it, checked against the deployment.
Verified
CSV export
Downloaded in a test, byte order mark checked, header compared with the columns the application shows.
Verified
Content publishing
An article published from the console and then fetched from the public site, then taken down and fetched again as a 404.
Verified
Contrast against WCAG AA
5,555 text nodes measured in a real browser across 40 routes.
Verified
Consent before analytics
A test watches the network and asserts no Google request is made before a visitor chooses.
Verified
Mobile layout at 375 px
Four pages checked for sideways overflow, and the mobile screenshot set captured at that width.
Verified
Deployment and database
The health endpoint reports managed Postgres and the real record counts from the deployed site.
Built, not yet verifiable
Email, thirteen wired messages
Rendered and checked through the file driver. Real delivery needs a provider key, which this build did not have. A welcome and a weekly summary are written and not wired, and the case study says so.
Built, not yet verifiable
Analytics event layer
Nine named events, a consent gate and a tag container slot. Nothing is active because no container id exists yet.
Blocked
The tracklyfitness.com domain
The domain is attached to the project and waiting. Its DNS is at a registrar this build could not sign in to, so the records could not be changed.
Blocked
Resend sending domain
No Resend session was available, so no API key and no domain authentication could be set up.
Blocked
Search Console, GA4, Tag Manager, Ads
The Google account signed in on this machine has no Analytics account, no Tag Manager account and no Search Console property. Creating them under the wrong identity would have been worse than leaving them.
Out of scope
Payments, wearable sync, food database
None of these is implemented and none is claimed. The feature pages say so in the panel on each one.
Out of scope
Medical or diagnostic features
Deliberately absent. The product records what a member enters and says so in the footer, the emails and the terms.
18
Screenshot library
47 screens, captured from the running application. 59 were taken in the last run, 51 at 1440 px and 8 at 375 px.