Leela Aravind.

Independent builder · Chester, England

Leela Aravind Karlapudi

I build practical web tools — and check that they actually work.

I'm an AI graduate and independent builder based in Chester. I turn ideas into scoped projects, direct AI-assisted development, and then investigate how the results actually behave. Most of what I know comes from building and testing my own products end to end.

Discuss a project View my work

Based in Chester, UK · Available for remote projects

Available for: small websites and landing pages · website testing and issue reports · AI workflow planning · site updates and fixes

What you can hire me for

Services

No fixed price list. Tell me the goal and I'll scope it.

Small websites and landing pages

Who it helps
Sole traders, freelancers and small businesses who need a clear, fast page that explains what they do.
You get
A responsive site or landing page, built and deployed, with the content structure agreed up front and one round of revisions.
I need from you
Your text and images (or a clear brief so I can draft them), plus access to wherever the domain is managed.
Limits
Small scope by design — a handful of pages. Not e-commerce platforms, custom dashboards or ongoing content operations.
Contact me for a scoped quote

Website testing and issue reports

Who it helps
Anyone who suspects their site has problems but wants evidence rather than opinions.
You get
A written report with reproducible steps, screenshots, expected vs actual behaviour, impact, and a suggested fix for each issue — plus an explicit list of what I did not test.
I need from you
The URL, and confirmation you own or are authorised to test the site.
Limits
Functional, content and layout checking in a real desktop browser. Not a security audit, not penetration testing, not load testing, and not a certified accessibility conformance statement.
Contact me for a scoped quote

AI workflow planning and setup

Who it helps
People who know AI could help with a repetitive task but don't know how to break it down or where to put the human checks.
You get
A written workflow: the task split into steps, which step each AI tool handles, where a human reviews, and what "done" looks like — plus a working setup of the parts we agree on.
I need from you
A walkthrough of how the task is done today, and examples of good and bad output.
Limits
Practical workflow design using existing tools. Not custom model training, not production ML systems, and not anything requiring your confidential data to leave your control.
Contact me for a scoped quote

Website updates and small fixes

Who it helps
Owners of an existing site who need changes made without hiring an agency.
You get
The agreed changes made, tested and deployed, with a short note on what changed and how I checked it.
I need from you
Access to the site or repository, and a clear list of what you want changed.
Limits
Work is capped at the scope we agree before I start. Rescues of badly broken systems get quoted separately after I look.
Contact me for a scoped quote

On request — Short-form video editing: basic edits from footage you supply. Ask me before assuming availability; I don't currently show public samples.

Selected work

Six case studies

These are my own independent projects, not paid client commissions.

CS1 — Website QA report (sample)

Sample deliverable
Problem
Anyone hiring a tester wants to see the actual output, not a promise.
What I did
Audited my own homepage in a real desktop browser: content consistency, heading structure, link destinations, colour-contrast spot checks and a skip-link markup check. Found three reproducible content defects, each reproduced twice with screenshot evidence — a pricing statement on a product card that contradicted the site's own funding disclosure, a status badge whose tooltip contradicted the completed work described on the same card, and a missing space in a paragraph.
How AI helped
AI-assisted browser automation drove the page inspection and drafted the report; I set the scope, judged which findings were real, and rewrote every conclusion the evidence didn't support.
Why it matters
The report states what wasn't tested as prominently as what was. That's the standard I'd apply to your site.
Evidence / limits
Chrome 152 on Windows at a real 1920×945 viewport. Automated mobile and tablet checks were not completed — the tooling couldn't resize the viewport, and the report says so rather than claiming emulated passes. Console errors are recorded as inconclusive, not passed.
Download the sample QA report (PDF, 7 pages) →

CS2 — search.itisyou: finding and stopping a runaway

Paused by design
Problem
A scheduled rebuild on my own search engine quietly ballooned into roughly 153 million database row writes — against a corpus of only about 2,000 documents.
What I did
Stopped the bleeding first: cleared the cron triggers and paused the crawl queues without deleting anything. Then diagnosed it from read-only analytics. Two faults compounded — every index version stored a complete copy of the corpus, and the rebuild trigger compared two counts whose gap could never close, so it rebuilt everything every 15 minutes forever. I rewrote the index so a build costs only what actually changed, added a lock, a daily write budget, and a refusal to run when that budget is red.
How AI helped
I used AI to work through hypotheses and read the analytics faster; the incident timeline, the decision to stop before investigating, and the budget design were mine.
Why it matters
Most of engineering is what you do when something is quietly wrong. Stop, preserve evidence, then fix the cause.
Evidence
Written up as an architecture decision record with the full timeline, plus free-tier budget and storage-monitoring documents. Scheduled runs remain deliberately off.

CS3 — clean.itisyou.app

Live
Homepage of clean.itisyou.app, a browser-based tool for cleaning PDF, CSV and Excel files
Problem
Messy PDFs, CSVs and spreadsheets need restructuring, but uploading a financial or personal file to a stranger's server is a real cost.
What I did
Built a browser-first restructuring tool covering all three formats, each with its own detector, parser, transform and export path running in a dedicated web worker. Repeated headers get stripped, ragged tables squared up, every change listed, anything uncertain held for review, and the original bytes never touched.
How AI helped
AI-assisted implementation across the format engines; I designed the pipeline, the change-log/undo model, and the confidence rules for what gets held back.
Latest work
A 320px reflow fix and adversarial test hardening.
Evidence
40 test files spanning unit, integration, golden-file and end-to-end tests. The privacy claim is structural, not marketing: it deploys as static assets with no server-side code path, so there is nowhere for a file to be uploaded to.
Visit clean.itisyou.app →

CS4 — Project Intelligence Agent

Private system
Problem
I work across a lot of repositories, and the honest answer to "what does this project actually do now?" was usually buried in stale documentation.
What I did
Built a local-first tool that makes a codebase answerable. Point it at a folder or Git repository and it works out what the project is, indexes code and documentation in a structure-aware way, then answers engineering questions with citations back to exact files and line ranges — and refuses to answer when the evidence isn't there. It flags where documentation and code have drifted apart, selects the right tests to run when something changes, and can attempt a bounded fix in an isolated copy, leaving my real working tree untouched and every change reversible.
How AI helped
This is the clearest example of what I actually do — designing where AI is allowed to act, and where it must stop and cite evidence instead.
Limits
Private system, runs entirely on my own machine, no public demo. Described here in outline only.
Evidence
Certified against a 26-point acceptance matrix with recorded evidence for each point, alongside unit, integration, contract, end-to-end and stress test suites.

CS5 — Quantum-assisted intrusion detection (DrunkQubit)

Research
Problem
Would a small quantum classifier beat an equivalent classical one on network intrusion detection — or just look impressive?
What I did
Built a hybrid pipeline: a classical autoencoder compressing traffic features, then a 4-qubit variational classifier, benchmarked against a classical "twin" head on the identical inputs. Pre-registered the prediction before running, then executed on real IBM quantum hardware.
How AI helped
AI-assisted implementation of the pipeline and analysis code; the experimental design, pre-registration and the decision to publish a negative result were mine.
The finding
A statistical tie, and the repository says so in its own words — "24 quantum parameters MATCH a classical head on the same 4-number diet", not "beat". The hardware also contradicted my pre-registered prediction, and I wrote up the confound rather than hiding it.
Why it matters
I'd rather report a null result accurately than sell a result that doesn't hold.
Evidence
Run on IBM's ibm_marrakesh backend, 183 seconds of billed QPU time across four recorded job IDs, 1,024 shots, three random seeds.

CS6 — Thraksha Guardian

Frozen prototype
Problem
Can a phone examine its own security, explain it in plain language, and act on it reversibly — with nothing leaving the device?
What I did
Built an Android proof of concept where a small language model runs entirely on-device and interprets findings, but never holds authority: it can only propose typed actions that a deterministic policy engine approves or refuses, and every action is snapshotted so it can be undone. Findings are graded by how they were established rather than reduced to a single score, and the audit log is hash-chained.
How AI helped
AI-assisted Android implementation; the separation between interpretation and authority was the design question I set out to answer.
Limits, stated plainly
Sideload-only with no Play review, all real-world evidence from a single device, no independent security certification, and no encrypted-traffic inspection. Deliberately frozen.
Evidence
A signed private-alpha build, phases 0–12.1 passing, with real-device evidence from a Samsung S20 FE on Android 13.

More projects

Independent work

Status is what the evidence supports today, not what was planned.

Live

  • Homepage of kits.itisyou.app listing browser-local developer tools
    kits.itisyou.app

    Browser-local developer utilities. 135 tools declared in a single canonical registry, 137 pages built, 70 unit test files and 19 browser test specs.

    kits.itisyou.app →
  • Homepage of qa.itisyou.app, a privacy-first QA workbench
    qa.itisyou.app

    A toolkit for testers: around 24 tools covering requirements, test-case design, coverage, boundary and pairwise analysis, and evidence building. Version 1.0.1 came out of a deliberate adversarial campaign against my own work that found and fixed 19 defects.

    qa.itisyou.app →
  • Homepage of space.itisyou.app, a calm landing page with no feeds or notifications
    space.itisyou.app

    A quiet corner of the web with no streaks, feeds or notifications. Verified live and checked at both desktop and 390px widths.

    space.itisyou.app →
  • Homepage of dev.itisyou.app, an evidence-based engineering playbook site
    dev.itisyou.app

    Cross-vendor engineering playbooks where evidence is attached to a specific revision and environment. Partially complete, and its own status document uses an explicit honesty ladder: "A successful build is not TESTED. A successful deploy is not VERIFIED."

    dev.itisyou.app →
  • itisyou.app

    The main network site.

    itisyou.app →

Built, not released

  • Bank Format Studio

    Offline Windows converter and validator for MT940, camt.053, CSV and XLSX bank formats. Installer built and checksummed; release held pending an anti-virus vendor review of a false positive.

  • PreflightQC

    Offline video-export spec validator with 12 platform presets. Packaged locally but explicitly not distributable, and never yet run end to end against real media.

  • products.itisyou

    E-commerce implementation with server-side pricing, payment-intent verification, idempotent webhooks and stock enforcement. Passes 994 tests including full checkout journeys in payment test mode. Never launched: no customers, no sales, no revenue.

  • plan.itisyou

    Planning workspace where every value is typed as actual, assumption or calculated, and provenance is enforced at compile time — putting AI in the calculation path is a type error.

Research and paused

  • AI-native phone OS

    A frozen architecture specification for a phone where agents propose and deterministic code authorises, with no inbound endpoints and egress denied by default. Setup-stage work only on real hardware; no OS exists.

  • itisyou-os

    An experimental x86_64 Rust operating system that boots under emulation. Research only — explicitly not for real hardware.

  • browse.itisyou

    A privacy-first desktop browser shell with AI off by default. Phase 1 of 4, paused. No downloadable browser.

  • Video production pipeline

    A local-first automated video pipeline using Remotion, FFmpeg and local text-to-speech, producing rendered drafts with per-scene quality frames for review. Nothing has been published or monetised.

  • Chaty

    A conversational-AI platform with a human-in-the-loop training review queue. Retired over hosting costs; kept as a case study.

Skills and approach

Grouped by outcome

Building

Small websites, landing pages, browser-based tools, and content or feature changes to sites that already exist.

Checking

Reproducible issue reports, user-journey walkthroughs, boundary and edge-case thinking, and being explicit about what was not tested.

Organising

Turning a vague goal into a scoped brief, breaking work into tasks, coordinating AI agents, and putting review checkpoints where they matter.

Investigating

Debugging, incident response, root-cause work, and reading evidence rather than trusting a status document.

Writing

Technical documentation, architecture decision records, and honest status reporting.

Tools I've used in my own projects

  • Python
  • JavaScript / TypeScript
  • React
  • Next.js
  • Node
  • FastAPI
  • Cloudflare Workers, Pages & D1
  • Supabase
  • Stripe (test mode)
  • Resend
  • SQLite
  • Playwright
  • Kotlin / Android
  • PyTorch
  • Qiskit
  • Remotion / FFmpeg

Used in my own projects, at the depth those projects needed. I direct and review AI-assisted development rather than claiming expert command of every framework here.

Process

How I work

  1. 1

    Understand your goal

    What outcome are you actually after, and who is it for?

  2. 2

    Agree a small, clear scope

    Written down, including what's out of scope, before anything starts.

  3. 3

    Build or investigate

    AI-assisted, reviewed by me at each step.

  4. 4

    Verify and show evidence

    You see what I checked, and what I didn't.

  5. 5

    Deliver and revise

    Handover plus the agreed round of revisions.

I use AI tools throughout my work and I review everything they produce. Anything you share with me may pass through those tools, so please don't send me material you can't share with a third-party AI service.

About

About me

I finished an MSc in Artificial Intelligence at Liverpool John Moores University in 2025, after a B.Tech in Computer Science Engineering in India. Since then most of my learning has come from building my own products and then trying to break them.

What I enjoy is understanding how a system actually works — spotting where it will fail, splitting a messy problem into tasks a machine can do reliably, and deciding where a person still has to look. I use AI heavily, but I don't hand it authority: the interesting design question is always where the evidence has to be checked.

That's the same habit I bring to client work. I'd rather tell you what I haven't verified than hand over something that looks finished.

Get in touch

Discuss a project

Tell me what you're trying to achieve and I'll tell you honestly whether I can help.

It helps if you include:

  • What you want built, fixed or checked
  • A link to your existing site, if you have one
  • Any deadline you're working to
  • Roughly what budget you have in mind

I'm a one-person operation, so replies may take a day or two. No guaranteed turnaround until we've agreed a scope.