AI/ML Engineer · Researcher · Islamabad, Pakistan

I build AI that survives contact with the real world.

Most models look great in a notebook. Mine have to keep working after they ship, with real users, messy data, and tools that don't always do what they're told. That's the thread running through a production AI platform at DevNeuron, an external federated learning research collaboration, and a tool I built to break AI agents on purpose before someone else does.

See the work Ask my site a question >_
GitHub LinkedIn Scholar Email
Portrait of Omer Bin Dawood, AI/ML engineer
BS Data Science · '27 📍 Islamabad, PK

01 / About

Reliability is a systems problem, not a model problem.

I'm from Sheikhupura, a small city in Punjab, and I now live in Islamabad. I didn't set out to do AI research. I set out to build things that actually worked, which turned out to be a much harder problem than any machine learning course lets on.

The gap between "the model works" and "the system is reliable" is where I've spent most of the last two years. It showed up first at DevNeuron, watching a conversational AI product handle real property leads instead of curated test cases. It showed up again studying how AI agents can share skills across clients that don't trust each other with their data. Both taught me the same lesson from different directions: an impressive benchmark number and a system you can actually depend on are not the same thing, and the difference is usually invisible until something breaks in production. That's the question I keep circling back to, in whatever form the current project poses it: what does it actually take to keep an AI system honest once it leaves the lab?

1+years in production AI engineering
2deployed AI products shipped
Strongdistinction-level CGPA, BS Data Science
Opensource AI security tool, Aginiti-RedTeam

02 / Experience

Two problems, two very different shapes

One is shipping a product real customers depend on. The other is building a marketplace from nothing. Most of what I know about reliability, I've learned by doing both at once.

AI Engineer

Jun 2025 to Present

DevNeuron · California, USA (Remote)

Day to day, I build full-stack, real-time, data-driven agent systems, rather than research prototypes that only have to work once, on stage. Direct customer feedback, not just a benchmark, is what actually shapes performance, compliance, and usability here.

  • I lead engineering on Leasap (leasap.com), a conversational AI platform for property management that's actually live, not a demo. It's handled 500+ property leads and 1,000+ AI-assisted conversations at roughly 2 seconds per response, running lead qualification, follow-ups, and tour scheduling without a human in the loop.
  • Watching that system run in production taught me where AI actually fails in the wild: rarely with an obviously wrong answer, more often with an output that looks reasonable while the system quietly mishandles a tool call, loses track of state, or drifts from what the user meant. That observation directly shaped how I later approached evaluation on Aginiti.
  • I also led development of Aginiti-RedTeam, an open-source, newer take on Aginiti (aginiti.com), DevNeuron's commercial agent-security product: it ranks attacks with an adaptive planner instead of probing blind, reaching the same findings with roughly 5x fewer requests than fixed-order testing (repo →).
  • On Aginiti itself, I built the evaluation pipelines behind it: 1,000+ adversarial agent evaluations at ~250ms median API latency, the infrastructure that turns "we think this agent is safe and compliant" into something you can actually measure.
  • Stack, for anyone curious: Python (FastAPI/Flask), PostgreSQL/Supabase, RAG pipelines, Docker on Railway and Render.

Co-Founder & AI/Software Engineer

Ongoing

QaamQaaj: AI Workforce Enablement Platform

The other half of my time goes to something I'm building from zero, for users the tech industry usually designs around rather than for.

  • QaamQaaj connects informal skilled workers in Pakistan with employers through web, voice, and chat. We've tested it with 70+ users across 200+ job listings, and I'm working through matching problems across 500+ records where the usual "clean structured data" assumptions mostly don't hold.
  • Because many of our users have limited literacy or limited smartphone access, I built CNIC/KYC verification with OCR and computer vision, and designed voice/IVR interfaces that don't assume someone can comfortably navigate a text-heavy app.

Cold Calling & Sales Expert

Dec 2023 to Mar 2025

Marketing Transformers · Sydney, NSW, Australia (Remote)

Before AI engineering, I spent time on the other side of the funnel, actually talking to the people a product is supposed to convert, which turned out to be a better education in what "the system works" means than any course.

  • I led targeted marketing campaigns using data analytics and CRM tools, lifting customer engagement by 20% and revenue by 15%.
  • I developed sales strategies that improved lead conversion rates by 10%. It's the same instinct for "what actually moves the number" that now shows up in how I evaluate AI systems instead of just benchmarking them.

03 / Education

The credentials, for the record

None of the above happened in a vacuum. Here's the academic side of it.

Bachelor of Science in Data Science

Expected Jan 2027

COMSATS University Islamabad · Islamabad, Pakistan

Academic standing

Distinction-level CGPA throughout the program; final semester in progress, with performance trending upward toward graduation.

Standardized tests

GRE: top percentile, with a perfect Quantitative score · English: advanced proficiency, IELTS-certified

Relevant coursework

Machine Learning, Artificial Intelligence, Statistics & Probability Theory, Linear Algebra, Differential Equations, Calculus & Analytical Geometry, Data Structures & Algorithms, Parallel & Distributed Computing, Information Security, Data Mining, Big Data Analytics

Before COMSATS

F.Sc. (Pre-Medical), Government Graduate College Sheikhupura (2020–2022) · Matriculation, Muslim Public Model High School (2018–2022)

04 / Research

ClientFedSkill

Trustworthy & Reliable AI Systems Federated Learning & Distributed Training AI Agent Evaluation, Security & Alignment Privacy-Preserving Machine Learning Multi-Agent & LLM Systems Client-Side Personalization

In collaboration with Dr. Sabiha Afroz and Dr. Ahmad Faraz Khan (Virginia Tech / Google), we're studying how AI agents can share reusable skills and reasoning patterns without centralizing personalization on a server that only ever sees partial information about each client's private tools and constraints.

My contribution has focused on the experimental side: comparing federation configurations, reviewing execution transcripts, and running the analysis that surfaces patterns aggregate results conceal. Two findings stood out. A recurring fallback mechanism, where a failed learned merge gets replaced by an unadapted, highest-reward peer solution, fired on roughly three of every four merges in analyzed runs. That's far more than the rare edge case it was designed to be. And a separate source of evaluation noise: a worker's own previously successful solution sometimes failed on an identical repeat, pointing to run-to-run nondeterminism as a distinct, compounding cause rather than a single merge bug.

A manuscript covering these findings is in preparation for submission to MLSys 2027.

~75% of analyzed merges hit the unadapted fallback path
19 comparable task families studied (6 improved, 7 unchanged, 6 regressed)
22.2pp task-level swing between configurations hidden by aggregate pass rates
66.7%→8.3% success-rate drop on one task family under a single configuration change

Aginiti-RedTeam: Adaptive Red-Teaming for AI Agents

May 2026 to Present

DevNeuron · open source

Most red-teaming tools test an AI agent one attack at a time, learning nothing from the last attempt. Before writing any code, I reviewed the published literature on adversarial testing and agent security, including IKEA, SECRET, the Interrogation Attack, and SPE-LLM, and built Aginiti-RedTeam around a different idea: a planner that treats the target as an evolving system with state and history, not a function to score once. It ranks every eligible attack from an 11-method catalog by expected information gain and diagnosed failure patterns, executes the top pick, logs what happened as Fact → Observation → Claim evidence, and repeats, reasoning turn by turn the way a human red-teamer would rather than following a script. Every finding is cross-checked against an independent, non-LLM disclosure oracle before it counts as confirmed.

11 attack methods in the catalog, 4 as full standalone attacks with published methodologies
10+ published papers reviewed before designing the benchmark
~5x fewer requests than fixed-order testing to reach the same findings
1,845 tests, fully offline, full suite runs in under 30 seconds

Validated against NVIDIA's garak: exact agreement on every directly comparable category, plus real findings a REST-only scanner can't see at all. It's an open-source, newer take on Aginiti, DevNeuron's commercial agent-security product.

View repository →

05 / Work

Selected work

Deployed

Leasap

The AI receptionist that never sleeps: qualifies leads, chases up the ones gone cold, and books tours over WhatsApp, running in production for a real property management business, not a pitch deck.

500+ leads1,000+ conversations~2s latency
leasap.com →
Open Source

Aginiti-RedTeam

Built to break AI agents on purpose, so a real attacker doesn't get the chance first. An adaptive planner accumulates evidence about a target and picks the next attack from an 11-method catalog, instead of treating every attempt as an isolated probe.

11 attack methods~5x fewer requests
View on GitHub →
Commercial

Aginiti

What Aginiti-RedTeam became once it needed to run at scale for paying customers: a commercial platform for stress-testing autonomous agents before they meet the real world.

1,000+ evaluations~250ms latency
aginiti.com →
Co-Founder

QaamQaaj

A workforce marketplace for Pakistan's informal skilled-labor economy, built for users who may never have used a job-search app before, with CNIC/KYC verification via computer vision built in.

70+ users200+ listings
Research

Healthcare AI & Computer Vision

Where I first learned that a model that performs well on a benchmark and a model you'd trust with a diagnosis are not the same claim. Computer vision workflows for early disease detection that fed into a peer-reviewed publication.

PythonTensorFlowOpenCV
Research Tool

GAIS: Graduate Admissions Intelligence System

I built this for myself while researching PhD programs, then kept expanding it: an AI-assisted admissions tool aggregating structured data with semantic retrieval, vector search (Supabase pgvector), and a Gemini-powered LLM interface on Next.js and FastAPI.

110+ universities3,000+ faculty profiles2,500+ admissions records359+ program deadlines

06 / Publications

Publications

Two published so far, with a third (the ClientFedSkill work above) in preparation. None of these are a substitute for an independent research program, which is exactly what a PhD is for. Full list on Google Scholar.

07 / Beyond the code

The résumé stuff, plus the parts that don't fit on one

Leadership & Activities

  • Lead, Computer Vision Circle
    COMSATS Research Club · Jan 2026 – May 2026

    Lead student research activities in computer vision and mentor peers working on applied AI problems.

  • Campus Director & Marketing Lead
    Hult Prize COMSATS · Jan 2024 – Apr 2025

    Led a 35-member team to revive Hult Prize COMSATS, securing funding, mentoring startups, and executing high-impact events.

    View credential →
  • Lead, Graphics Team
    C3 Club, COMSATS

    Managed design and visual communication for club events and initiatives.

    View credential →
  • Graphic Design Lead
    Google Developer Group (GDSC), COMSATS

    Contributed to innovation, collaboration, and the club's digital presence in the campus tech community.

    View credential →

Certifications & Awards

  • Micro1 Certified Data Scientist
    View credential →
  • Meta Certified Data Analyst
    Coursera
    View credential →
  • Introduction to Cryptocurrency
    Simplilearn
  • Best Debater of the College
    Government Graduate College Sheikhupura

Languages

EnglishUrduPunjabi

When I'm not at a keyboard

BadmintonGym & FitnessReading & ResearchAutomation & Robotics

08 / Skills

The toolbox

For anyone doing due diligence before a call.

Machine Learning

PyTorch, TensorFlow, scikit-learn, NumPy, Pandas, OpenCV, Deep Learning, NLP, Computer Vision, Statistical Analysis

Federated & Privacy-Preserving ML

Federated Learning, Differential Privacy, Byzantine-Resilient Aggregation, Client-Side Personalization, Distributed Training, Model Heterogeneity

AI Agents & LLMs

Large Language Models, Retrieval-Augmented Generation, Agentic Systems, Tool Use & Function Calling, Prompt Engineering, Semantic Retrieval, Vector Embeddings, Agent Evaluation, AI Security & Red Teaming

Programming & Systems

Python, SQL, JavaScript, TypeScript, Java, C++, FastAPI, Flask, Node.js, NestJS, React, Next.js

Data & Infrastructure

PostgreSQL, MySQL, MongoDB, Supabase, pgvector, Docker, AWS, Railway, Render, Linux, n8n, Git, REST APIs

Data Science & BI

R, Jupyter, SciPy, Apache Spark, Hadoop, Kafka, Power BI, Tableau, Looker Studio, Snowflake, BigQuery, Apache Airflow, ETL Pipelines, Data Mining, Data Warehousing

09 / Ask

Ask my site a question

Reading is optional. This is a small scripted assistant, entirely client-side, just pattern matching over the content on this page. Try "who is Omer", "what do you work on", "clientfedskill", "aginiti", "leasap", "qaamqaaj", "gais", "skills", "education", "cold calling", "certifications", "leadership", "publications", "hobbies", "contact", or "hire".

guest@omerbindawood: ~
>

10 / Contact

Let's talk.

If you're a professor, a research collaborator, or someone with an interesting engineering problem, I'd like to hear about it. Based in Islamabad, Pakistan, and happy to work across time zones.

omerbindawood@gmail.com LinkedIn GitHub Scholar