CS student · AI / machine learning / computer vision

ZohaPasha

I build systems, then take them apart to find out what they're actually doing.

Lahore, Pakistan
CGPA
3.86

of 4.00

Honours
6

Dean's ×4 · Rector's ×2

Next
2026

AI/ML roles · grad study

01

What I work on

01

Computer Vision

Getting a model to recover detail that isn't obviously in the pixels.

This is where my final-year work sits. I implemented a published DSen2-based super-resolution architecture for Sentinel-2 satellite imagery — channel attention, high-frequency filtering, and the full preprocessing and evaluation pipeline. Working through someone else's architecture line by line taught me more about why it works than reading about it did.

02

Language Models

Building things with them — and checking what they are actually doing.

I built a debugger that surfaces a model's behaviour per token: confidence, entropy, attention weights, hallucination risk. Then I benchmarked whether its stated confidence could be trusted, and it couldn't — roughly 32% calibration error on what it claimed against 15% on its own token probabilities. I'm interested in the whole stack around these models, from wiring them into an application to understanding why they behave the way they do.

03

Deep Learning

From the gradient up, not from the API down.

Neural networks, CNNs, ensembles, reinforcement learning — through coursework, Andrew Ng's specialization, and building things until they worked. I'd rather implement a method and watch it fail than take a library's word for it. It is the layer underneath everything else here.

02

Selected work

012026

LLM tooling

LLM Reasoning Debugger

A local web app that opens up a language model's reasoning token by token — confidence, entropy, attention weights, and hallucination risk in a terminal-style interface.

  • Served Qwen2.5-1.5B-Instruct behind a FastAPI backend, capturing per-token logits and attention.
  • Built a 28-question benchmark and scored it with AUROC and expected calibration error.
  • Found the model's stated confidence was badly calibrated (~32% ECE) against its own token probabilities (~15% ECE).
~32%
ECE, stated confidence
~15%
ECE, token probability
  • Python
  • FastAPI
  • Qwen2.5-1.5B

022025

Paper implementation

Sentinel-2 Super-Resolution

An implementation of a published DSen2-based super-resolution architecture for Sentinel-2 satellite imagery, reproduced end to end to understand how channel attention and high-frequency enhancement actually behave.

  • Implemented the paper's channel attention mechanism and high-pass frequency filtering.
  • Ran the full pipeline myself: dataset preprocessing, training, and quantitative evaluation.
  • This is a reproduction of existing published work, not an original architecture.
  • Python
  • PyTorch
  • CNNs

032025

Game / agent behaviour

AI Murder Mystery

A story-mode murder mystery where the NPCs decide for themselves — behaviour trees driving patrol, chase, and attack states, with language-model dialogue on top.

  • Designed behaviour trees for NPC patrol, chase, and attack state transitions.
  • Wired pathfinding into level geometry so pursuit reads as deliberate rather than scripted.
  • Integrated an LLM for reactive dialogue during the investigation.
  • Unreal Engine 5
  • C++
  • Blueprints

042025

Full-stack platform

Airline Management System

A booking, scheduling, and administration platform built on a relational database with a Next.js frontend.

  • Modelled flights, bookings, and schedules in SQL Server.
  • Built the booking and admin flows end to end.
  • Next.js
  • Tailwind CSS
  • SQL Server

052024

Desktop app

FAST Learning System

A past-paper search platform for FAST students, with an AI chatbot and Supabase handling auth and storage.

  • Built the desktop client in WPF and C#.
  • Indexed past papers for fast retrieval and added a chatbot for student queries.
  • WPF
  • C#
  • Supabase

062024

Desktop app

Social Networking System

A desktop social app implementing friend requests, messaging, and user profiles.

  • Built friend-request graphs, profile state, and messaging in Qt.
  • C++
  • Qt
03

About

I'm a final-year Computer Science student at FAST-NUCES Lahore, sitting at a 3.86 CGPA with four Dean's List and two Rector's List placements.

Most of what I build comes from the same impulse: I want to see inside the thing. That's how I ended up writing a debugger that reads an LLM's confidence token by token, reproducing a super-resolution architecture from a paper to understand why channel attention works, and wiring behaviour trees into game NPCs to watch them make decisions.

I'm heading into AI and machine learning work. Engineering roles and graduate study are both on the table and I'm keeping both open while I finish my degree — computer vision, language models, and deep learning are the areas I want to keep building in.

Education & experience

  1. 2023 — nowEducation

    BS Computer Science

    FAST-NUCES Lahore

    • CGPA 3.86 / 4.00
    • Dean's List ×4 · Rector's List ×2
    • Coursework: Applied Machine Learning
  2. 2025Experience

    Developer Intern

    Clear Wave Information Technologies

    • Frontend features in Next.js and Tailwind CSS
    • ERP implementation and testing support
  3. 2024Experience

    Frontend Development Intern

    OSOL Technologies

    • Responsive UI components in HTML, CSS, and JavaScript
    • REST API integration support
  4. 2026Certification

    Machine Learning Specialization

    DeepLearning.AI — Andrew Ng

    • Supervised learning, neural networks, ensembles
    • Unsupervised learning, recommenders, RL

Tools

AI / ML

PyTorchTensorFlowscikit-learnCNNsNeural NetworksLangChainLangGraphNumPypandas

Languages

PythonC++C#JavaScriptAssembly (8088)

Web

Next.jsTailwind CSSFastAPIASP.NETWPF

Tools

Unreal Engine 5GitSQL ServerSupabase
04

Contact

Working on something in AI, machine learning, or computer vision? I’d like to hear about it.

Elsewhere

LinkedInGitHub

Lahore, Pakistan