Zarmeen Tauseef

Academic

Verbatim text

Zarmeen Tauseef
+923349993618, zarmeenta@gmail.com
..:::...
F-17 /2, Islamabad
Linkedln: htt~s:[Lwww.linkedin.com[in[zarmeen
Education
Bachelor of Science (Artificial Intelligence)
Major:
~
eep Reinforcement Learning, Computer Vision, Generative Al, Digital Image Processing,
rtificial Neural Networks, Machine Learning, Natural Language Processing, MLOps
)UperNova School, Islamabad
~ Levels (Mathematics, Economics, Accounting)
Merryland International School, Abu Dhabi
GCSE (Mathematics, Physics, Chemistry, Biology, ICT, English)
Projects
-;no/ Project: Salon360: Salon Management and Virtual Hair Try-On [Python, React, OpenCV, Flask, Node.js]
I\ salon management web application with Al-driven virtual hair try-on. Includes analyzing hair type, length, and face
,hape for feasible stvling classifications while ensuring real-time svnc with salon operations.
,emester Proiects:
Real-time Safety Analysis at Worksites [ YOLOvll, Flask, and CV]:
Built a system to provide individual and scene-wide safety ratings with color-coded alerts.
Hair Colour Transformation:
:::arried out a comparative analysis of using CycleGAN, Style Transfer, Diffusion and UNIT for the task.
ormula One Knowledge Graph and Web Linkage:
Used Graph DB, Owlready2, and SPARQL to create a queryable knowledge graph.
Multimodal Generative Al Story App:
Developed a React Native app integrating Hugging Face API models for audio, image, and text generation.
~ir Xonix Game:
Used x8086 Assembly Language to create a 2D recreation of the classic game Air Xonix.
Multimodal Meme Sentiment Analysis:
mplemented custom end-to-end architecture from scratch using PyTorch.
Work Experience
~ssociate Al Automations Engineer, VECTOR Inc., Islamabad
Dec 2024 - Present
Building and deploying voice Al agents with enterprise workflows and system integration
Deep Learning Engineer (Part-time), VECTOR Inc., Islamabad
Feb 2024 - May 2024
Researched and worked on vFit (VTON), cut inference time by 1 s for real-time try-ons
)oftware Engineer (Machine Learning) Intern, VECTOR Inc., Islamabad
Jul 2023 - Aug 2023
reaching Assistant and Lab Demonstrator, FAST NUCES, Islamabad
Feb 2023 - Present
Skills & Tools
Professional Skills
Problem Solving, Teamwork, Interpersonal Skills, Communication, Adaptability, Critical Thinking
fechnical Skills
Python, C++, C, Assembly, Bootstrap, JavaScript, NodeJS, SQL, HTML, CSS, Git, MySQL, Visual
:itudio, PyCharm, React, React Native, Protege, GraphDB, VS Code, Jupyter Notebook,
TensorFlow PvTorch sci kit-learn OWLReadv2 Keras Numpy & Pandas
Achievements
Bronze Medal (Fall 2022), Dean's Honor List (Fall 2022, Spring 2022, Fall 2023, Spring 2023, Fall 2024)
Runner-up of Parwaz-e-Takhayyul ldeathon hosted by GDGoC at FAST NUCES
Trainings I Certification
l\ttended Flutter workshop by Google 1/0 Extended at FAST NUCES, Al for Everyone - DeepLearning.AI Certificate
Activities
Host for IMPACT-ISTAN at NaSCon '24 Hosts Team Vice Head (GDSC 2023-2024), Content Team Lead (FAIS 2023-2024)
'nterests
ormula 1, Painting
FAST NUCES ISLAMABAD CAMPUS
FAST SCHOOL OF
COMPUTING
FINAL YEAR PROJECTS
FAST NUCES ISLAMABAD CAMPUS
Agential Chitchat
This project aims to develop a mechanism that ensures a consistent communication protocol
across agents trained at different times. This includes designing a system where new agents can
align their communication with previously trained agents and ensuring that agents can learn both
independently and collaboratively while preserving consistency in message representations. To
address this challenge, we propose a training framework that incorporates historical agents as part
of the learning process of newly introduced agents. This approach enables: Adaptive Message
Encoding: Agents are encouraged to learn representations that remain compatible with earlier
communication structures. Cross-Epoch Alignment: By incorporating previously trained agents
during training, we ensure that new agents inherit and refine established communication
strategies. This research will contribute to MARL by introducing a robust communication
mechanism that allows for seamless collaboration between agents trained at different times. The
findings will be particularly useful for applications in:
•
Autonomous Robotics: Ensuring long-term adaptability of multi-agent robotic systems.
•
Decentralized Al Systems: Enhancing interoperability in Al-driven multi-agent
environments.
•
Game Al & Simulations: Improving coordination strategies in Al agents deployed in
competitive and cooperative games .
I l,.klm A ll
211 0316
•
.,..-
Supervis or O•. Ahmau 0'n
o ... m s o ood ,,. 035'
AG ENTIAL-CH ITCH AT
o k>f:> a
trD.ntl~orablo and ger><0ro.lizablo corn~cnticn •1.1atom
1n mulU-~nL &nVif'"onrn•nta. a..--blln,g language am--~·~• and
n<:loptioo-. YJlthoVt •Y>O<;tlf'OJ-ng ~nt
pol'CU $ truotur()s. ens.uri.-.g
$CO:.IO.l:>ll14..1 a,,-.d lo-.Le.'Of'>ernbmtu.
VVo rkf'lovvr of' Algorit hm
T lmeHne
" "'"o>•t <><:<U<'OM"'..,....
• ""'""'•c•..,.oo_, . .,
l
--~
+!Hi
+
• l!Ji.!i:[ n•
.
~., .. ,_ . ...
~ .... ,. ... ___ ,,_
. · ~""'" ....... "'~~·" ......
,.....,_ ___ ..
•
.,,...,,.. .
,~d---" "" '-",..u.,. ..
. : "~:::::-~~'' "' '""''-"" .. " ......
---~
l
.n . ..,,,. . ., • .,..,.,,
T ool s. and T uohnologluw
<$> -WRLLc;>I: .4
Technology Used:
Python, pytorch, Manim
pettingZoo, Visual studio
Supervisor Name:
Dr. Ahmad Din
Group Members:
Qasim Saeed (211 - 0352)
Hakim Ali (211 - 0316)
FAST NUCES ISLAMABAD CAMPUS
Auto Morph
We explored training different diffusion model architectures in order to modify existing drivers and
synthetically augment them so they appear as 'drowsy'. These were the following approaches used:
•
Mask-based Text-conditioned Augmentation: Given a mask and specific keywords
describing the desired drowsy effect on the masked region, a diffusion model would denoise
the masked region into the desired drowsy effect. (E.g. If you mask a person's mouth and
gave the text prompt of 'yawning', it should change the person's mouth into a yawn while
maintaining good quality of the image.)
•
Pose-guided Augmentation: Given an image and a corresponding pose (DW Pose), alter the
image to that respective pose's structure. (E.g. If you give an image of a person looking
straight and pose keypoints of the pose looking right, the newly generated image should be
of that person looking right.)
As such, these techniques were used to generate augmented datasets which were used to enhance
the accuracy of driver safety detection algorithms (Mainly detecting if the person is yawning or has
his eyes-closed)
Key Words: Generative Al, Diffusion Models, lnpainting, Drowsiness/Yawning detection, Driver Safety
Detection Systems
AUTOMORPH
:t,"....,~~
.M
Al Driven synthetic driver ddtd Enr1Chmen1
~
rj
tor enhanc€'d detect10'1 systems
"~,,.....,-e
•
TEAM
Syed Haider Naqvi (201-0816)
Sarim Aeyzaz (211-0328)
MOTIVATION
SUPERVISOR
Dr. Asif Naeem
Detecting driving behavior like yawning for an Al model is tough due to the
limited datasets publicly available. AutoMorph fills this gap by using normal
images of drivers and generates their distracted behavior counterparts
by taking in additional inputs such as a mask, text, or pose key points by
using diffusion models and the latest techniques of Generative Al.
ARCHITECTURE
K•Y P<Hnt Em~ddln&
AUTOMORPH MODEL
···w
.. -
~ ~
Image Input• Mask
--
-..
/
0-'Mn&U· Net
Concat• naud
ultlmodal lnpuu
Auamented lm•a•
Text Input
Technology Used:
Python, PyTorch, AWS, Gradio, Diffusers
Supervisor Name:
Dr. Asif Naeem
Group Members:
Sarim Aeyzaz (i21 - 0328)
Syed Haider Naqvi (i20 - 0816)
FAST NUCES ISLAMABAD CAMPUS
TherapEase -Al-powered Therapist Assistant
TherapEase is an Al-powered therapist assistant designed to enhance autism assessment and
therapy using a 3D Digital Twin model. It features Therapist View, allowing professionals to monitor
patient progress, track session data, analyze Automated Diagnostic Decision Support (ADDS) scores,
and receive real-time emotion detection insights. The Patient/Guardian

AI enrichment