AI builder & testerToronto

Hello, I'm

Jaskaran

I build things with AI, then test them until they break.

Scroll

I test by buildingwhat I do

I build real things with AI tools, then try to break them. Building shows me what a tool can do. Breaking it shows me where it stops.

My fitness app, Taakat, went live after 634 automated checks and 170 break-in tests on its database. Two of its bugs only showed up on my own iPhone, which is why I test on the real device every time.

When something breaks, I report exactly what I saw, with screenshots, and keep going until the cause is found.

How I workfour habits

  1. Rules first. Before a build starts, I write down how the AI has to work with me: ask before building, one change at a time, tell me when my idea is a bad one.
  2. Real devices decide. A version is done when it works on an actual phone, not when it looks right on a laptop.
  3. Fix the cause, not the test. If a check fails, the check stays as it is.
  4. Say no. I dropped calorie counting from plate photos, because wrong numbers are worse than no numbers.

Backgroundthe short version

I have a diploma in computer programming, but I never worked as a programmer and I don't write code today. I use AI tools the way most people will, without knowing how to work around their gaps. In a tester, that is useful.

Based in
Toronto, Canada
Looking for
AI tool testing roles
Makes
Apps, video ads, voice agents
Works with
Claude, Higgsfield, Google Flow, CapCut
Also used
Vapi, n8n, Supabase, Gemini
Education
Diploma in computer programming
Code by hand
None. Everything here was built with AI.

Selected builds

Five things I've made with AI tools.

Each one is labelled with the state it's actually in, including the one that isn't finished.

02.

Unique Solutions of Media, a demo site

A demo website for a real family business, a film production house in New Delhi. One page with a showreel, a filterable list of projects, services, training and a contact form. It is a demo: the contact details are placeholders and the form doesn't send yet.

Demo Open the site ↗
03.

Ivy, a voice receptionist

An AI receptionist for a made-up Toronto hair salon. She answers by voice, checks the calendar and books the appointment. One browser call went through end to end: right day, right time, on the calendar and in the sheet.

  • Testing turned up three bugs
  • She booked every appointment into 2024. The date tool answered too slowly, so she guessed the year.
  • She announced failures a second before the booking succeeded.
  • Booking details arrived as one lump of text, so the start time came through empty. No error, just wrong bookings.

Not done: it has never taken a real phone call, she promises transfers she can't make, and the calendar check takes 7 to 9 seconds. It isn't running today.

Paused Built with Vapi and n8n
04.

BOLT, ads for a made-up energy drink

BOLT is an imaginary energy drink brand I built for this portfolio. These are two ads for it, each made with a different AI video tool.

The dragon ad, 43 seconds. Made in Higgsfield with Seedance 2.5. The main character is my AI avatar. One flaw I left in: the model garbled the background text on a billboard and a chalkboard.
The stadium ad, 10 seconds. Made in Google Flow with Omni Flash 1.1.
Finished Higgsfield  ·  Google Flow
05.

Kit swap, an AI car video

An 18-second clip: a car drops into a parking lot, then cycles through four body kits on a hand snap. I generated the footage in Higgsfield. The edit, effects and music were done with AI, under my direction.

Finished Higgsfield

Say hello

Hiring a tester, or want to see one of these builds up close? Email me.

jaskaransingh2291@gmail.com
Jaskaran Singh  ·  Toronto, Canada © 2026

Build 01  ·  Fitness app

Taakat

A workout and food app I built with AI, without writing the code myself.

$0a month to run, on free plans
634automated checks before a release
170break-in tests on the database
3weeks from idea to live, about

What it isਤਾਕਤ, strength in Punjabi

Taakat is a personal app with two tabs. Food shows the calories you have left today, with protein, carbs and fat. Training logs a workout set by set and fills in what you lifted last time.

It installs on a phone's Home Screen from one web address, and works on iPhone, Android and laptop. Sign-up needs an invite code, so it stays private.

Inside
860 foods, 162 exercises, 8 ready-made workout plans
AI tools
Describe it (type what you ate) and Scan a label (photo of a nutrition label)
Built with
Claude, Supabase, GitHub Pages, Google Gemini
Version
4.6.4, October 2026
Today screen: 533 calories left, with protein, carbs and fat
TodayCalories left, and the three macros under it.
Describe it screen: a typed meal split into two foods with calories
Describe itType a meal. The AI splits it into foods, marked as an estimate.
Training screen: day 3 of 6, Legs A
TrainingThe ring shows which day of the plan is next.
Workout in progress with sets ticked off and a timer
WorkoutTick each set. A timer runs at the top.
Exercise page with a progress chart
ProgressA chart and personal bests for every exercise.

My rolethe AI wrote the code

I ran the project. I decided what to build and what to skip, picked the design, did the setup in GitHub and the database myself, and tested every version on my iPhone before it went live. The AI never had my passwords or secret keys.

Before any code, I wrote rules for how the AI had to work with me:

  1. Before building anything, ask me the questions you need answered. Ask one question at a time.
  2. Recommend one build approach that fits a non-coder, and list any trade-offs. Wait for my approval before starting.
  3. Build in small stages. Do one change at a time, then tell me how to test it before moving on.
  4. If something I ask for is a bad idea or won't work well, tell me honestly and suggest a better option.
  5. Never lose our data. Before any change that could affect saved logs, warn me and explain how to back up first.

How I tested itfour layers

I couldn't check the code myself, so I needed proof that things worked. A change only shipped when every test passed. If a test failed, we fixed the cause, not the test.

207

Robot tests

They open the app in a hidden browser and use it like a person: log in, log food, start a workout. They run on a fake copy of the database, so real accounts are never touched.

257

Maths and logic checks

Calorie targets, protein goals, weekly weight change, which workout day comes next, and how AI answers get cleaned up.

170

Break-in tests

Checks that act like an attacker. They try to read other people's rows, change someone else's data and sign up without a code. I ran them on the real database after every database change.

Real

My own phone

Every version, in Safari and as a Home Screen app. Some bugs only show up on a real phone, and this layer caught them.

Bugs I chased down

The text box that disappeared

Problem
The Describe it box showed on my laptop but was missing on my iPhone.
Cause
My phone's ad blocker. The box's internal name started with "ad-", so it was hidden as an advert.
Fix
Renamed it, and added a test that fails if anything is ever named like an ad again.

The AI that never answered

Problem
The AI kept loading, then gave up.
Cause
The model names written into the code had changed, and the bigger model spent too long "thinking".
Fix
The server now asks Google which models exist, picks a fast one, and reports exactly what went wrong.

Buttons that needed two taps

Problem
On iPhone, tapping a button while typing often did nothing the first time.
Cause
The tap closed the keyboard first. The page jumped, so the tap landed in the wrong spot.
Fix
The app handles the tap the moment your finger lifts. This bug never showed up on a laptop.

What I said no toand why

We discussed a third AI tool: snap a photo of your plate and get the calories. I rejected it. Guessing portions from a photo is too inaccurate, and wrong numbers are worse than no numbers.

The two AI tools that did ship are labelled every time. An entry says "AI estimate" or "from label (AI)", so you always know where a number came from. The app checks every number the AI returns and drops anything missing or unrealistic.

Three looksone kept

For each redesign the AI drew two or three full mockups side by side, and I picked one. Nothing got built until I said yes.

September

Hybrid

Dark, glassy cards. It worked, but felt flat and looked AI-made.

Early October

Scoreboard

Bold, sporty, all capitals. It had character, but felt loud.

Mid-October

Nightglow

Near-black with a soft mint glow and light, thin numbers. This is the one I kept.

What I'd do differently

  1. Set up a demo account from day one, so anyone could try the app safely.
  2. Measure how accurate the AI is from the start, not at the end.
  3. Use a test copy of the app from the first version. I only added one once other people started using it.

See it live

The app is private, so the link opens the login screen. If you'd like to try it, ask me for a demo login.

Open Taakat Back to all builds
Jaskaran Singh  ·  Toronto, Canada © 2026