Hi, I'm Qat, an AI and machine-learning engineer based in Lahore, Pakistan.
Most AI just gives an answer and hopes it's right. I build the part that checks the answer, says how sure it is, and shows the proof. That is the kind of system companies rely on to run AI safely in the real world.
99.6% correct on real customer email, deciding live in production right now.
0accurate on live customer email
0to decide each one, faster than a blink
0messages handled a day, on one machine
0products shipped: 3 to download, 1 on Google Play
Walkthrough · 5 minutes
See it in my own words
Five minutes on the systems I've built, the projects, and the numbers behind them. No jargon required.
What I've built
Real products, with real numbers
Not side projects. Things that run in production, that you can download today, or that are live in an app store. Every claim below links to the proof.
Running in production(handling real traffic right now)
Email classification system
At Data443, an AI system that reads real customers' email as it arrives and sorts it (safe, spam, or dangerous) without slowing delivery down even slightly.
99.6%
correct, live in production
A mail server can't wait around, so the system has just a few thousandths of a second to decide on each message. Instead of asking one slow, expensive AI model every single time, it uses fast, cheap checks to clear the easy 99.7% of mail instantly, and only calls on the heavy AI for the genuinely tricky ones. Whenever it does get one wrong, that mistake is automatically fed back in so the system keeps learning from real mail instead of going stale.
Under the hood Gemma 3 model, fine-tuned (PEFT/LoRA), served on vLLM · 2.66 ms average decision, under 12 ms at p99 · ~2M messages/day on a single A100 GPU · false-positive/negative reports feed the retraining loop.
Published & downloadable(anyone can install and use these today)
Avow
A tool for software teams that builds code and then double-checks its own work before it trusts it.
284
of its own tests passing
Most tools stop the moment the tests pass. Avow goes a step further and asks the harder question: "were those tests actually good enough to believe?" Then it gives one honest confidence score instead of a plain green checkmark, and openly flags anything it isn't sure about rather than pretending everything is fine.
Under the hood 8,625 lines of Python · hold-out splits, mutation & property testing, reference oracles, differential fuzzing · open source (Apache-2.0).
Verbatim
A shared memory for teams. It remembers what was decided and promised across your meetings, chats, and project discussions, so nothing slips through the cracks.
519
of its own tests passing
It pulls together meetings, Slack threads, and pull-request discussions into one clear picture of who committed to what, by when. It catches contradictions, writes your daily stand-up update for you, and can run entirely on your own computer when privacy matters.
Under the hood 23,555 lines of Python · one state graph of commitments/decisions/blockers · MCP server for other AI agents · hosted or fully local models · MIT.
Vaikora
A safety checkpoint for AI assistants. It checks every action an AI wants to take against the rules before it happens, and keeps a tamper-proof record.
<500ms
to approve or block an action
As companies let AI do more on their behalf, someone has to make sure it stays inside the lines. Vaikora sits in front of the AI, allows or blocks each action in real time, and seals every decision into a locked record that an outside auditor can replay and verify independently, without having to trust anyone's word for it.
Under the hood adapters for OpenAI, Anthropic, Gemini & OpenRouter (HTTP, MCP, agent-to-agent) · verdicts: allow / log / constrain / block · SHA-256 append-only audit chain · SDKs on PyPI & npm · MIT.
Live in people's hands(through app-store review and into real users' phones)
Susanoo
A phone app that shows you, in real time, which companies your apps are quietly talking to in the background.
Live
on the Google Play Store
Your apps constantly send data to companies you've never heard of. Susanoo shows you exactly who, as it happens. Everything is worked out right on your phone, and nothing is ever sent anywhere else. That makes the privacy structural rather than a promise you have to take on trust.
Under the hood Kotlin · reads DNS & TLS SNI through a local VpnService · on-device classifier attributes each hostname · classifier is open source so it can be checked.
Convaa
A shared AI workspace where a whole team can watch the same AI conversation happen live, together, on the web and on iPhone.
Web + iOS
both built and shipped
Everyone on the team sees the AI's answer appear word by word at the same time, can leave comments, turn those comments into assigned tasks, and even stop the AI mid-answer while keeping whatever already arrived. The iPhone app keeps working even with no signal, so a message written on the subway still sends the moment you're back online.
Under the hood Web: React, Supabase realtime, Postgres row-level security · iOS: 14,613 lines of Swift, OAuth handoff, offline send queue · tested with XCUITest driving the real screens.
Tools
What I reach for
Chosen because a system had to actually run, not because a name was on a trend list.
Python · TypeScript · Swift · Kotlin · Java · C · SQL
Every number here can be checked.
If I say 99.6% accuracy or 519 tests, you should be able to go and see for yourself. Three of these are one click to install, one is on Google Play, and the code is public wherever the licence allows. I'd rather be checked than believed.