Hirak Banerjee · Quality report · Ontario, Canada

Senior Quality Engineer · 14 years · banking, payments, insurance & wealth
St. Thomas, Ontario (London area) · remote across Canada, hybrid in southwestern Ontario · Canadian citizen, no sponsorship needed

I build AI into how software gets testedand I verify what AI builds.

Hirak Banerjee
Fig. 1Hirak Banerjee, St. Thomas, Ontario

For fourteen years I've owned quality for financial systems where a bug costs real money: bulk payments at one of Canada's largest banks today, and insurance and wealth platforms for Manulife and Canada Life before that. Now I pair that discipline with LLMs. I build AI test tooling my team uses every release, and I ship AI-orchestrated products that I measure like production systems.

Experience
14yrs

QE in regulated finance

At the bank
3

AI-built testing tools in use

LLM runs
499

logged and measured, not estimated

DayDrop installs
5.5K

175 countries · 4.6★ from 81 ratings · as of Sep 28, 2026

Awards
3

innovation and project awards

§1At the bank
2021 – present

Tools I built with AI, used every release

Bulk payments on a Big-5 Canadian bank's platform: Kafka event flows, database and API validation, and a regression cycle that is now almost entirely automated, with scheduled nightly regression, smoke suites and CI/CD gate runs in Jenkins, built largely with AI. These three tools came out of that work and each was built in days, not months.

T-01

Requirements in, test cases out, written by an LLM

Paste a user story or requirements page. It returns structured test cases as a reviewer workbook or a file ready to upload into test management. It also converts hand-edited sheets.

● LLM-powered
GitHub Copilot

T-02

One-click payment trace with audit evidence

Follows a single payment from the payment network through the database to its Kafka event, then writes the tracking record and a step-by-step evidence document with screenshots. Runs on a laptop or from Jenkins.

● Built with AI
Copilot · Claude Opus

T-03

Payment test-data generator

Generates valid bulk payment test files across transaction types. One configuration per environment, so a scenario runs everywhere. No dependencies.

● Built with AI
Copilot · Claude Opus

Described at a high level on purpose. Client systems, code and data stay with the client.

§2My own work
2025 – present

Products I orchestrate with AI, and measure

SwiftGen AI

Multi-agent iOS app generator · solo · prepared for Y Combinator

You describe an app in one sentence. SwiftGen plans it, writes native SwiftUI, compiles it, repairs what breaks, launches it in the iOS Simulator and streams it to your browser, where you keep changing it by chat.

I built it alone: 315 commits in three and a half months, about 90,000 lines. The interesting part isn't the generation. It's everything that stops a model's output from reaching a user unchecked.

  1. 01Classify. Intent and complexity decide the route and the model.
  2. 02Generate. A router across Claude Sonnet and Opus, Gemini, Grok and GPT, with fallbacks and truncation recovery.
  3. 03Pre-validate. About 970 injected rules and 193 fix patterns, no LLM call.gate
  4. 04Build and heal. Three repair attempts, escalating to a stronger model.gate
  5. 05Verify. Launch, screenshot, process check, crash-log scan.gate
  6. 06Learn. 600+ repairs logged; repeat fixes promoted into the validator.
Table 1. Results from the pipeline's own logs
MeasureResultBasis
New app, idea to running66%333 generations · 71% over the last 100
Change by chat83%166 modifications
Failed builds rescued~46%by the healer loop
Time to a running app2.4mmedian, successful runs

Measured January to April 2026 on that generation of models (Claude Sonnet 4.6 and Opus 4.6, GPT-4.1, Gemini 2.5). The harness doesn't care which model it tests, so I can re-run it on current ones.

Workout tracker generated by SwiftGen
"a workout tracker"
Personal finance dashboard generated by SwiftGen
"a personal finance app"
Food delivery app generated by SwiftGen
"a food delivery app"
Shopping app generated by SwiftGen
"an online shop"
Project board app generated by SwiftGen
"a project board"

Fig. 2. Simulator screenshots taken by the pipeline's own verify step, Jan 2026.

DayDrop

iOS app · 5,500+ installs · 175 countries · 4.6★ from 81 ratings · as of Sep 28, 2026

Countdown widgets with AI-generated backgrounds, Dynamic Island Live Activities and iCloud sync. I built it with an AI coding agent and tested it like production: committed XCUITest regression suites, checks on real devices, and verification against sources the AI didn't write.

For example, before the lunar-calendar release every date from 2026 to 2028 was checked against the Hong Kong Observatory's official tables: 1,078 days, zero mismatches. The listing and screenshots ship in 16 languages through an automated pipeline.

DayDrop countdown list
Countdowns
DayDrop event detail with AI background
AI background

Stackwrite

Developer reference · written by an 8-agent pipeline

A reference site on developer-tool pricing and limits, produced by agents on a schedule. The part I care about is the gates. The writer can't add a fact the researcher didn't source. A fact-checker re-fetches every source and defaults to hold. A reviewer opens the built page on desktop and mobile. Nothing ships without a pass.

research → write → visuals → fact-check → seo/geo → review → publish → index

Underlined steps can stop the run. 50+ pages now surface in Bing's top 7 for the questions AI assistants ask.

SiteAuditAI

Web app · live · one-click website audit with paid reports

Paste a URL and get a plain-English report on accessibility (WCAG 2.1 AA via axe-core), security headers, SEO, performance and cross-browser rendering, with the risk explained by Claude and a Stripe-gated PDF. Next.js and TypeScript on Supabase; the scanner runs real browsers through Playwright.

Tested the way I'd test a bank's checkout: 11 Playwright end-to-end specs cover the scan flow, purchase flow, subscriber and upsell states, the monthly cap, pricing and accessibility; 9 API test suites cover feature gating, checkout and scan routes. Entitlements and payments are verified, not assumed.

url → crawl → axe-core + headers + SEO + vitals → screenshots → Claude summary → risk score → Stripe → PDF

Bold steps are gated by tests. The second domain is the plain-language funnel into the same scanner.

OSS-1

testgen-evals

An evaluation harness for LLM-generated test cases: deterministic metrics (schema, acceptance-criteria coverage, duplicates, untestable steps) plus an LLM-as-judge rubric with a position-bias check. In the committed run, tag-level coverage read 100% while the judge scored 9 of 52 cases 2/5 on completeness. MIT.

GitHub ↗

OSS-2

RadioWave

A macOS menu-bar app that turns Claude Code's live activity into synthesized radio sound, through Claude Code hooks. MIT.

GitHub ↗

OSS-3

llms-txt-init

One command that scaffolds an llms.txt and an AI-crawler-aware robots.txt for any site.

GitHub ↗

OSS-4

tiktok-produce

Screen recording in, short video out: voiceover, synced captions and a 1080×1920 render in one command.

GitHub ↗

OSS-5

ConsensusAI archived · Jan 2026

A multi-LLM debate engine: GPT, Claude, Gemini and Grok argue three rounds to a verdict. Shelved before launch; kept public for the test suite — Playwright/Puppeteer UAT flows on desktop and five phone viewports, a CI runner with smoke/mobile/desktop/security modes, and unit tests for moderation and API resilience. Node.js, MIT.

GitHub ↗

§3Method

AI output gets verified, not trusted

1

Fix the system, not the output

Never hand-patch what a model wrote. Fix the prompt, the validator or the healer, then re-test simple, medium and complex cases.

2

A green build proves nothing

Launch it, tap it, screenshot it, read the crash log.

3

Measure, then claim

Every number here comes from logs or live data, with its date and the models it ran on.

4

Use an outside oracle

Check AI-written logic against something it didn't write: official tables, the spec, a second system.

§4Career

Fourteen years in financial software

2021 – now

Senior Software Quality Engineer, AI-Augmented Test Automation

Tata Consultancy Services · for a Big-5 Canadian bank · Payments

  • Quality for bulk payments across Kafka, Oracle, MongoDB, SQL Server and mainframe; automated nearly the whole regression and release cycle, with scheduled regression, smoke and CI/CD gate runs in Jenkins.
  • Built the three AI testing tools in §1.
  • Java, Selenium, Cucumber BDD and REST Assured frameworks on the bank's enterprise test-automation standards, plus XCUITest mobile suites.

2012 – 2021

Senior Quality Engineer and Test Lead

Cognizant · Insurance & Wealth · Manulife, Canada Life and a major US insurer

  • Led testing in an onshore–offshore model: estimates, strategy, sign-off, triage and mentoring.
  • Automation in Selenium/Java, TestNG, Cucumber BDD, Robot Framework and UFT with Jenkins CI; API, ETL and data testing.
  • Took automation proofs of concept to break-even and worked on mainframe modernization.

★ Best Innovation 2016 · ★ Best Innovation 2014 · ★ Best Project 2014

2008 – 2012

B.Tech, Civil & Transportation Engineering

National Institute of Technology, Agartala

Testing: Java, Selenium, Cucumber BDD, TestNG, REST Assured, XCUITest, Appium, Jenkins, Docker. Payments & data: Apache Kafka, Oracle, SQL Server, MongoDB, mainframe. AI & LLM engineering: multi-model orchestration and routing (Claude, GPT, Gemini, Grok APIs), schema-enforced output, validation and self-healing loops, evaluation harnesses, prompt engineering, token and cost tracking, Claude Code, GitHub Copilot, MCP. Building: Swift and SwiftUI, Python, FastAPI, Next.js, Docker, Cloudflare.

Side B · off the clock

Songs, a film and a channelmade after the day job, with the same eye for what feels off.

Singles
2

written, produced, mixed

Short film
1

festival submission

Subscribers
1.8K

in two months

One video
257K

views

TracklistLiner notes

A1

Unkaheen single cover

Unkaheen

Single · Hindi · written, composed, produced, mixed and mastered

The first release. Produced in FL Studio and GarageBand, from first idea to master.

Spotify ↗

A2

Sach single cover

Sach

Single · Hindi · written, composed, produced, mixed and mastered

The latest release, built the same way in FL Studio and GarageBand.

Spotify ↗

B1

Short
film

The short

Writer · director · editor · producer

Made end to end, from pre-production through the edit in Sony Vegas Pro to the final grade and sound mix, and submitted to a film festival.

Not online yet

B2

Football
Memes
AI

FootballMemesAI

YouTube · creator · editor

Short-form football content, scripted, edited and published. 1,800 subscribers in its first two months, and one video reached 257,000 views.

YouTube ↗

Music FL Studio · GarageBand · composition · mixing · mastering

Picture Sony Vegas Pro · DaVinci Resolve · CapCut · colour · sound design

Words scripts · lyrics · storyboards

Day job Side A, the engineer →

Contact

hirak.b@hotmail.com

Canadian citizen · St. Thomas, Ontario · remote across Canada or hybrid in southwestern Ontario · no sponsorship needed

LinkedIn ↗ GitHub ↗ X ↗ testgen-evals ↗ YouTube ↗ Music ↗