Work / 01

Live

2026

Tooldo

Nine everyday file tools, built so that privacy is architecture instead of a policy page.

Role
Solo. Product design, design system, architecture, and build.
Year
2026
Timeline
Aug 2026
Team
Solo build. Every decision here, from the palette to the deploy target, was mine to make and mine to get wrong.
Stack
Astro, React, TypeScript, Tailwind CSS, pdf-lib, Mediabunny, Cloudflare
Tooldo's landing page: the headline Fix any file. Keep it private. above the invoice tool.

01 — The 30-second version

The 30-second version

Problem
Free file tools all want the same thing first, your file on their server.
What I did
I designed and built nine everyday file tools that run entirely in your browser, under one brand and one design system.
Result
Live at tooldo.online. Zero files uploaded and zero outbound requests from any tool, 437 automated tests, 21 days from empty repository to live domain.

Read the full story

02 — Overview

Nine everyday tools, and none of them can see your file.

This did not start as a product idea. It started as the same complaints, from different people, over and over.

The first one was photos. A friend gets sent a picture from someone’s iPhone, it arrives on their Windows laptop as a .heic file, and nothing on the machine will open it. I watched people hit this more than once, and the fix is always the same small sad ritual. Search for a converter. Land on a page covered in ads. Upload the photo to a stranger. Wait. Download it back.

Then it was QR codes. People kept needing one for something, a menu, an event, a form, and the free generators either stamp a watermark on it or want an account first.

Then it was invoices. A few friends who run their own businesses were either paying a monthly subscription to send an invoice, or building them by hand in Word every single time.

Three different complaints on the surface. Underneath, all the same shape: a small boring task you need done once, which the internet has made either expensive or sketchy. Nobody wants a relationship with an invoice tool. They want one invoice, now, and then they want to leave.

And when I actually looked at what those sites do with your file, the pattern got worse. That photo, which might just as easily be a passport scan or a signed contract, gets uploaded to a company you had not heard of ninety seconds earlier, and sits on their disk for as long as they feel like keeping it.

So the question I started with was not “what tools should I build”. It was: what would these look like if the file simply never left your device?

Tooldo is nine of those everyday tools, PDF, images, video, data and generators, under one brand and one design system. None of them can see your file. Not because we promise not to look, but because there is no server anywhere in the product that could receive it. Everything runs inside your own browser tab, works on your own machine, and hands the result straight back to you. Close the tab and it is gone.

The tool list is not arbitrary either. HEIC conversion, the QR generator and the invoice builder are all in there because I watched somebody I know get stuck on exactly that thing.

I designed and built it myself, in the open, MIT licensed.

Not because we promise not to look, but because there is no server anywhere in the product that could receive it.

03 — The core tension

The decision that makes this product worth trusting is the same decision that makes it hard to build.

The decision that makes this product worth trusting is the same decision that makes it hard to build.

No server means the browser does everything. Video encoding, image codecs, PDF manipulation, all of it has to run on whatever device a stranger happens to be holding, in whatever browser they happen to have. There is no beefy machine in a data centre to fall back on. Every hard problem stays hard, and it stays on the client.

Underneath that sat a second tension, and it was more of a design problem than an engineering one. Nine tools is not one product by default. They have genuinely different shapes. Some take a file you drop in. One takes a form you fill out. One takes text you type. One takes nothing at all and just needs to know what city you care about. Left alone, that becomes nine small websites wearing the same logo.

And a third, which is the one I kept circling back to. A utility tool has to be memorable enough that someone bookmarks it instead of re-googling it next month, and calm enough that you can actually get work done inside it. Those two goals pull in opposite directions. Loud brands are tiring to work in. Quiet ones are forgettable.

So the brief I gave myself was: make the privacy claim structurally true, make nine different shapes feel like one thing, and be loud in exactly the places that do not cost the user anything.

04 — What we learned

The fear is real, it is specific, and it is about one question.

Three inputs, and I want to be straight about what each one is worth.

The first was just paying attention to the people around me, which is where the tool list came from. It is not a formal study and I am not going to dress it up as one, but hearing the same three complaints from different people, unprompted, is a real signal about what is actually annoying in ordinary life.

The second was a competitive teardown. Converting the same HEIC file, compressing the same video, merging the same two PDFs across the tools that rank for those searches.

The third was going and reading what people had already written publicly about this exact problem, in support forums, in security reporting, and in one government advisory.

The fear is real, it is specific, and it is about one question. People are not vaguely uneasy. They want to know where the file physically goes. A user on Adobe’s own community forum, asking about their PDF export service:

“I work for a legal firm and most of the documents uploaded to convert to PDF will be highly confidential. what are the risks in using this service? In the modern age of cybercrime I am concerned this information can be accessable? where are the files actually stored when the conversion takes place?” Adobe Community, January 2017

The reply he got from another user is the part that stuck with me:

“Review your company’s security policy in the light of this. It is not likely that it is permissible.”

That is the whole product brief, sitting in a support thread. Someone with a real job and a real deadline, who cannot use the obvious tool, because nobody can tell him where his file ends up.

The uncertainty does not slow people down, it stops them. Another user on the same forum, six years later:

“Since I am not sure if ADobe uses the data for its private usage through cloud, I am not able to use the PDF converter for document conversion.” Adobe Community, January 2023

He did not shop around for a better privacy policy. He just did not use it. That is the actual cost of an unanswerable question, and it is worth noticing that this is Adobe, a company with an enormous amount of trust to spend. If the doubt survives that brand, it is not a branding problem.

The fear turns out to be well calibrated. In July 2024, Cybernews found two online PDF services, PDF Pro and Help PDF, leaking 89,062 user-uploaded files through unsecured cloud storage. Passports, driving licences, certificates, contracts. Attempts to contact the operators went unanswered, and people kept uploading documents the entire time it was open. Then in March 2025, the FBI’s Denver field office put out a public warning that free file converter tools were being used to deliver malware and to scrape uploaded files for social security numbers, dates of birth and phone numbers.

So the average person’s instinct about these sites is not paranoia. It is roughly accurate, and the tools have earned it.

Which produced the reframe the whole product came from. Every one of these services has a privacy page, and most of them say reasonable things. Files deleted after an hour, we never look at your data. But every one of those sentences is a promise about future behaviour, and promises change when a company gets acquired or runs out of money or misconfigures a storage bucket. Nobody was making a claim you could actually check.

If the file physically never leaves the device, the promise stops being a promise and becomes a property of how the thing is built. It cannot be quietly revised in a changelog, and a breach cannot expose files that were never collected.

That led straight to a principle I leaned on constantly: let people verify it themselves rather than asking them to extend trust. The privacy page tells you to open your browser’s network panel and watch, and says plainly that you will see the page load and nothing leave. The source is public for the same reason. It answers that legal firm question directly, and the answer is that there is nowhere for the file to be stored.

Two smaller findings that shaped the interface. People come to do one thing and leave, so there is no onboarding worth building and no account worth having, which makes the empty state the front door rather than a fallback screen. And the category has trained everyone to expect a trap, ads, forced sign-ups, a paywall halfway through the download, so the landing page ended up answering “what is the catch” section by section instead of listing features.

Sources: Adobe Community, 2017 · Adobe Community, 2023 · Cybernews, July 2024 · FBI Denver, March 2025

05 — Decisions made, and why

Both were caught by checking the thing rather than trusting the plan.

  1. No backend, at all.

    Not a small one, not for the heavy tools. The moment one tool has a server, the promise on the front page needs an asterisk, and the whole proposition falls over. This also made hosting effectively free at any scale and shrank the attack surface to almost nothing, since there is no database, no auth, and no API to attack. The cost is real and I accepted it deliberately: heavy tools now depend on the visitor’s own device.

  2. Logic never lives in the interface.

    Every tool’s actual work sits in a plain, typed function file with no framework anywhere near it, and the React component on top is a thin shell that calls it. This sounds like a code cleanliness thing. It is actually what let me test all nine tools properly, because pure functions can be tested exhaustively against nasty edge cases and malformed files, and interfaces cannot.

  3. One canonical way to do each thing.

    One dropzone, one download helper, one toast, one empty state component. Never a second variant. This is the single decision that most made nine different tools feel like one product, more than the palette did.

  4. A registry as the only source of truth.

    Every tool is declared once in one file, and the nav, the landing grid, the sitemap and the routes all read from it. Adding a tool is one edit. It also meant the product could never ship a link to a page that did not exist yet, because unbuilt tools are marked as such and the nav knows.

  5. I killed a tool that worked.

    Background Remover shipped as tool three. It passed every automated check, including a test that downloaded and ran the real machine learning model, plus manual checks in a real browser. Then it failed for an actual visitor, repeatedly, on two unrelated networks. The cause turned out to be the model host’s own CDN caching an error at some edge locations, so the identical request to the identical file succeeded from one place and failed from another at the same moment. I shipped a retry with backoff. It did not help, because the bad cache sits there for as long as it sits there, and no amount of retrying rides that out.

    The fix that would have actually worked was hosting the 42MB model file myself. I chose not to. The entire premise of the suite is small self contained tools that keep working, and this was the only tool in the product that needed a service I did not operate. That dependency was the real problem, and the CDN was just the specific way it showed up. So I removed it and the suite went from ten tools to nine. Every tool in the product now makes zero outbound requests, which lets the privacy claim be stated plainly instead of qualified.

  6. The obvious library was the wrong answer twice.

    For video, I built the whole thing on ffmpeg compiled to WebAssembly, which is what everyone reaches for. It built, it typechecked, its tests passed. Then I checked a real production build against the actual deploy target and found the core file was over 30MB, against a hard 25MB per file limit on my host. It could never have deployed. I rebuilt the engine on the browser’s own video APIs instead, which took the whole thing from a 30MB binary to a few hundred kilobytes. For the data tool, the library named in my own spec turned out to have a stale distribution, last published to the registry in 2022. I wrote the parser by hand instead, character by character, so quoted commas and newlines inside fields do not break it the way a naive split does.

    Both of those cost me time, and both were caught by checking the thing rather than trusting the plan. That is the habit I would want anyone reading this to take away.

  7. The same habit paid off again on the landing page, months later in project time.

    The hero plays short screen recordings of the tools actually working. They came out of the recorder at roughly 6.5 Mbps, which is close to Blu-ray bitrate, for footage that is almost entirely flat colour and text. Five clips added up to 104MB sitting behind a hero that renders at 734 pixels wide, on a site whose entire build is 19MB. Re-encoding them at quality 90 took the set to 25MB with no change in resolution at all, nothing downscaled. I compared the two versions side by side before applying it, because the automatic setting visibly softened the small type in the invoice clip, and legible small type is the entire reason those recordings exist.

  8. Offline as proof, not as a feature.

    The site had been advertising an Install button through its manifest for weeks with nothing behind it, which meant anyone who installed it got a blank window the first time they opened it without signal. I wrote the service worker by hand, because the recommended plugin does not support the framework version this runs on and I would rather own forty lines than wire an untested dependency into the build. But the real reason to do it at all was that a tool which keeps working with the network physically switched off cannot possibly be uploading your files. Offline stopped being a nice extra and became the most direct demonstration the product has.

  9. A bug that existed only because of who I built this for.

    The invoice tool draws its PDF with a standard font that can encode western European text and nothing else. Type a naira sign into it and the whole document fails, and because that error had no specific handling, it reached the person as “the invoice didn’t generate, try again”, which is advice that can never work. The same characters fail identically every time, on a document somebody needs to send today.

    I only found it because I tested with a Nigerian bank detail, which is not an edge case here. The friends who wanted an invoice generator in the first place are Nigerian business owners, so a naira sign in the payment block is the most ordinary thing that field will ever hold. Currency symbols now substitute for their codes, which matches a decision the project had already made for the totals column, and anything genuinely undrawable names the character rather than asking you to try again forever.

    The general point is not about fonts. A default that quietly assumes Latin text will pass every test written by someone who only types Latin text.

  10. Honest copy, including where it is inconvenient.

    A tool you have never opened still needs a connection the first time, because its engine has not downloaded yet. So the copy says “the tools you’ve opened keep working”, not “everything works offline”. There is a test that fails if anyone ever softens that into the bigger claim. Same instinct elsewhere: the image tool tells you plainly that you cannot really compress a PNG without resizing it or converting it, instead of shipping a button that saves two percent and calls it compression.

06 — The design system beneath

Loud where it delights, quiet where it works.

The palette is called paper, ink and signal. A warm cream paper, a near black ink, and exactly one hot vermilion. Light first, warm throughout, because the whole category is cold blue and grey and I wanted the thing to feel like a well made paper object rather than a dashboard.

Accessibility changed that palette twice, and both changes stuck.

The vermilion clears contrast requirements as a surface, for focus rings and borders and progress bars, but white text sitting on it only reaches 3.6 to 1, under the 4.5 to 1 that small text needs. So a filled button could never carry a white label at that value, no matter how much I wanted it to. Rather than fudge it, the system grew a second, deeper step used specifically for fills with content on them, which gets white to 4.7 to 1. The hot value kept the jobs it was always right for. Separately, the faintest ink tone in the system was originally reading at 3.2 to 1, and hints and captions are text people actually need, so it got darkened until the palest text on the site still passes.

The design language is neo brutalism, chunky ink borders, hard shadows with zero blur, pill shapes, and a signature interaction where the shadow lifts on hover and the element snaps down onto its own shadow on press.

The rule that kept it usable is what I called the restraint rule. Full brutalism on the surfaces that are meant to delight, which is buttons, dropzones, cards, empty states, and the landing page. Deliberately calmer on the surfaces where real work happens, which is option panels, tables and long forms. Lighter borders, no hard shadows. Loud where it delights, quiet where it works. Without that rule, the option panel on the image tool would have been exhausting to use, and I know that because the first version was.

There is a similar line drawn for glass. The nav is frosted glass, and glass is also the right treatment for illustrations, like the folder inside a dropzone. The line is chrome versus artwork. Anything you operate stays brutalist, because a control needs a real edge and a focus ring. Anything purely drawn can be soft and blurred, because it is not a control and never needs a hit target. Glass artwork inside a brutalist frame is the combination. A glass button is not.

Nothing in any component holds a raw value. No hex codes, no pixel values, no durations. Colours, spacing, radii, type sizes, shadows, easings and durations all come from one token file. If a token is missing, the rule is to add it there, never to inline it. That is what makes a theme change a one file edit rather than an archaeology project.

Motion runs in three tiers, small interactions everywhere, transitions between states, and one or two big scroll moments on the landing page. Every non essential animation is gated behind reduced motion with an instant fallback, and that is enforced by a whole test profile that runs the entire suite with the preference on. The mobile version of the tools grid is a sticky card deck where each card pins under the nav and the next rides up over it, shrinking and dimming as it gets covered. Worth noting the dim is a brightness filter rather than opacity, because fading a card with opacity also fades its ink border and hard shadow, and you end up seeing the next card straight through the edges of the one on top. That was a real bug I caught by watching it on a phone, not by reading the code.

The last thing the system taught me was about optical versus measured spacing, and it came out of the landing page. The gaps between the rows in one section measured identically, and they were identical, but they read as uneven. The reason was that the divider line sat at the top edge of each row rather than centred in the gap, so the whitespace above a line was forty pixels and the space below it was seventeen. Every gap was a lopsided pair rather than one even band. Boxes measuring equal does not mean the page looks equal, and the page is right.

Colour

  • Paper#F4F0E7
  • Paper dim#EBE6DA
  • Ink#16130D
  • Ink soft#57534A
  • Signal#FF3B14
  • Signal fill#DE2E00

Type

  • Clash DisplayDisplay
  • InterBody and interface
  • JetBrains MonoCode and labels

07 — The build

Every tool's work lives in a plain typed function, and the interface is a thin shell that calls it.

Astro and React, in TypeScript, styled with Tailwind from one token file. Every tool’s work lives in a plain typed function, and the interface is a thin shell that calls it. There is no backend: PDFs, images and video are processed on the visitor’s own device, with pdf-lib, browser-side image codecs and the browser’s own video APIs. A hand-written service worker keeps the tools you’ve opened working offline. 437 automated tests run against desktop, mobile and reduced motion, and every one of the 124 commits passed typecheck, lint, tests and a production build before it went out.

08 — What was designed

Twelve surfaces owned end to end, from the token layer through to production deploy.

PDF Toolbox · Image Converter · Video Compressor · Invoice Generator · Screenshot Beautifier · CSV and JSON Viewer · Timezone Finder · QR Code Studio · Text Toolbox, plus the landing page, the privacy promise, and a living design system page. Twelve surfaces owned end to end, from the token layer through to production deploy.

09 — What it moved

"Zero files uploaded" is not a statistic that could drift next quarter.

The guarantees

files uploaded, ever, by construction
0
outbound network requests from any tool
0
accounts, ads, or rate limits
0
tools shipped, every one fully client side
9

The craft

Lighthouse accessibility, best practices and SEO, measured on the live site
100
automated tests, unit and browser
437
device profiles every test runs against, desktop, mobile, and reduced motion
3
architecture decisions written up as ADRs
17
engine test files covering pure logic, separate from the interface
19
contrast built into the palette itself, which twice forced a colour change rather than an exception
AA
to make the entire site work with the network switched off
3MB

The process

from empty repository to live domain
21 days
commits, every one passing typecheck, lint, tests and a production build
124
working, shipped, fully tested tool deleted on purpose
1

The domain went live this month, so there is no traffic story yet and I am not going to invent one.

What there is instead is a set of numbers about how it was built, and a few that I would argue are worth more than traffic anyway, because they are structural rather than measured. “Zero files uploaded” is not a statistic that could drift next quarter. There is no code path that could make it go up.

What I would measure next

Now that it is live, three things would tell me whether the thesis holds. Time from landing on a tool to having the finished file, which should stay under about thirty seconds for a single file job. The share of visits that arrive direct or by bookmark rather than through search, which is the real signal that someone chose it rather than settled for it. And the share of sessions that use the sample file, because that tells me whether the empty state is doing the job I designed it to do.

10 — Looking back

Knowing what a feature costs the whole, not just what it costs to build.

The thing I keep coming back to is the tool I deleted.

Background Remover was finished. It was tested, it was live, and it was the most impressive thing in the suite from a distance, because running a real machine learning model in someone’s browser is a good trick. When it started failing for a real person, my first instinct was the one I think most people have, which was to fix it in place. So I shipped a retry, and I was quite pleased with the retry, and it did not work.

Sitting with why it did not work is what actually taught me something. The retry was treating a structural problem as a temporary one. The tool was not broken, it was dependent, and it was the only thing in the entire product that needed something I did not own. I could have rescued it by taking on infrastructure, hosting a 42MB model file myself, versioning it, checking its integrity. And I would have been carrying that weight forever, for one tool, in a product whose whole argument is that it is nine small things that do not need anything from anyone.

Deleting work that functions is genuinely uncomfortable, and it is much harder than debugging. But the suite got better the moment it went from ten tools to nine, because the sentence on the front page stopped needing an asterisk. I think that is the most useful instinct I picked up here, and it is a product instinct rather than a craft one: knowing what a feature costs the whole, not just what it costs to build.

11 — Credits

Credits

Product design, design system, architecture and build
Emmanuel Onugwu

Next project

ShopAm

Products · Mobile · 2025

Keyboard shortcuts

CtrlK
Open the command palette
T
Switch between light and dark
G
Show the 12-column grid
?
Show these shortcuts
Esc
Close whatever is open