CS PhD researcher at the University of Windsor with three years of industry engineering behind
it. I work across LLM fine-tuning, reinforcement learning agents, evaluation, and
production ML — and I've shipped models that run on 80,000+ requests a day.
Multi-agent reinforcement learning for automated fault discovery, with reward
functions designed to push agents into unexplored territory.
algorithmsPPO · ICM · LLM-guided policies
faults found+29% over baseline
exploration+128% request diversity
Taking a model from notebook to live traffic — deployment, latency budgets,
monitoring, and the client conversation that goes with it.
latency10s → <1s per request
throughput80,000+ / day
ownershipdesign → deploy → evaluate
The unglamorous half: pipelines, cleaning, integration, CI/CD, and dashboards
that people actually make decisions with.
stackPython · SQL · Pandas · Tableau · Power BI
automationRPA + API workflows, CI/CD
impact+30% reach · −10% manual effort
Every number here comes from a shipped project or a published result.
01
How I work
Research to product
I've done both sides — publishing at ICSE and TOSEM, and shipping a fine-tuned LLM into a
live logistics pipeline. Neither is a detour from the other.
Measurement first
My research is about how you tell whether an AI system is correct. That habit follows me
into every project: define the metric before the model.
Full ownership
Model design, training, deployment, evaluation, and the client conversation. I've run
projects end to end and led junior engineers through them.
Teaching as testing
I run LLM and prompt engineering workshops for Master's students. Explaining a method to
someone new is the fastest way to find where it breaks.
02
Projects
Measured results
Grounded Judge Gate
Evaluation · label-free
A multi-agent code judge can produce claims, verify them, and report no errors while having
no basis for its verdict at all. Two measurements read straight from the pipeline's own
logs — no ground truth needed — catch that case, and a gate withholds those comparisons
instead of guessing.
gated_accuracy20.7% → 36.9%
coverage_kept50.6% of comparisons
gainpositive in all 27 cells
WiML @ NeurIPS 2026 · under review
Parcel Verification at Scale
Production LLM
Replaced regex and Google Maps matching with a fine-tuned LLM that resolves ambiguous and
incomplete delivery addresses. Owned it end to end: model design, deployment, client
handover, evaluation.
latency10s → <1s / parcel
throughput80,000+ / day
improvement75%+
PARL
Multi-agent RL
An LLM-driven multi-agent reinforcement learning tool for automated API fault discovery.
Diversified reward functions push agents toward unexplored request space instead of
re-testing what already passes.
Predicts whether generated code is correct without reference tests or execution — at a
fraction of the cost of running an LLM as a judge. Scores deterministically by combining
LSTM and GNN embeddings with transformer architectures, and introduces the
Alignment Score, a metric that classifies generated code against data labels.
approachreference-free, no execution
modelsLSTM · GNN · Transformers
contributionAlignment Score metric
venueICSE 2027 · under review
Code released after review
Thoth
IoT systems testing
End-to-end testing framework for consumer IoT, validated on real health-monitoring
deployments. Measures recovery and cascading failure behaviour, not just whether a device
responds.
Scalable pipelines at Robiquity integrating RPA tools, Python services, and cloud
infrastructure to move data between business systems — with junior engineers to bring
along on the way.
manual_effort−10% in AR processing
reliabilityerror rates down, fewer interventions
roletech lead + mentor
03
Research
06 papers
TMLRIn preparation
When Is a Multi-Agent Code Judge Actually Grounded? Identifying What Its Evidence Must Satisfy, and How to Measure It
Extended journal version of the WiML abstract below. Sets out the conditions a multi-agent
judge's evidence has to meet before its verdict counts as grounded — independence from the
answer under review, and difference between the candidates being compared — and proposes
measures for checking whether it does.
In preparationAly, S. R., Assaf, H., Kobti, Z.
WiML @ NeurIPS2026
When Is a Multi-Agent Code Judge Actually Grounded? Two Label-Free Measurements, and a Judge That Declines to Guess
Ran a published multi-agent verification pipeline unmodified on two code-judging
benchmarks, across 799 paired comparisons and two Qwen3 judges. It parses, verifies, and
reports no errors — while calling the two solutions equally good on most comparisons and
landing far below the same model asked directly. Two quantities computed from the
pipeline's own logs explain it, with no ground truth involved: how often the proposer
asks identical questions of both solutions, and how readily the checker agrees. Gating on
the first, before the checker runs at all, lets the pipeline answer only where it has
something to go on.
Thoth: A Lightweight Framework for End-to-End Consumer IoT Rapid Testing
The first metrics-driven framework for evaluating performance, reliability, recovery,
security, and cascading failures across a whole consumer IoT setup.
Splits hallucination checking across a Solver, a Proposer, and a Checker, then withholds the
original answer from the Checker so it can't rubber-stamp its own reasoning. Deliberate
information asymmetry as a fix for confirmation bias in LLM judges — and the framework my
WiML submission runs unmodified on code.
Trains a coder model to critique rather than only to answer, rewarding it when its
judgment of a solution matches the ground-truth label. Swapping a fifth of the RL data
for critique data beats RL-only baselines — and the gains carry over to general reasoning.
Asks whether agents can reason about code semantics without running it. Structured
"semi-formal" prompting forces explicit premises and traced execution paths, acting as a
certificate the agent can't skip steps in — with clear gains on patch equivalence and
fault localisation.
Argues that stacking LLM evaluators on LLM evaluators converges on internally consistent
but ungrounded verdicts. Judges agree with each other while drifting away from human
raters — which is the whole problem my own work is circling.
Fine-tuned VLMs score well in-distribution and fall apart under covariate shift, suggesting
training doesn't actually install the reasoning rule. The fix compiles the rules into a
symbolic circuit and leaves the model to do perception.
Chen, Vergari, Zhao
05
Experience
Research Assistant — University of Windsor
Sep 2023 — Present
Built a deterministic evaluation framework for LLM code responses using LSTM and GNN embeddings with transformer architectures.
Introduced the Alignment Score, a new metric that classifies generated code via data labels.
Developed two label-free measurements that detect when a multi-agent code judge has no basis for its verdict, and a gate that lifts pipeline accuracy from 20.7% to 36.9% while still answering half of all comparisons.
Designed PARL, an LLM-based multi-agent RL tool for automated API fault discovery — 29% more faults and 128% more diverse requests than baseline.
Designed and validated Thoth on real health-monitoring setups: 2.5s fault detection, sub-second recovery, under 1% energy overhead.
Research Project Mentor — University of Windsor
May 2024 — Present
Ran hands-on workshops on LLMs, prompt engineering, and software engineering practice for Master's students.
Led interactive sessions on prompt engineering, fine-tuning, and SE workflows, taking students from theory to working implementations.
Mentored graduate researchers through experimental design and evaluation of AI-driven software systems.
AI Consultant, R&D — BlinkTech Solutions Inc.
Sep 2024 — Mar 2025
Owned the full project lifecycle: model design, development, deployment, client collaboration, and performance evaluation.
Automated verification of unresolved parcels with LLMs, handling the ambiguous and incomplete addresses that regex and Maps matching couldn't.
Cut verification from 10s to under 1s per parcel — a 75%+ improvement enabling real-time processing of 80,000+ parcels daily.
Fine-tuned foundational LLMs in PyTorch and TensorFlow for specialised logistics use cases.
Intelligent Automation Engineer — Robiquity
Jun 2022 — Jul 2023 · Cairo
Built scalable automation pipelines integrating RPA tools, Python, and cloud services across business systems.
Led and mentored junior engineers on collaboration, code quality, and delivery.
Reduced error rates and manual intervention through end-to-end process automation, cutting manual effort in accounts receivable by 10%.
Data Analyst — The American University in Cairo
Jan 2022 — May 2022 · Cairo
Collected, cleaned, and analysed social media datasets with Python (Pandas, NumPy) and SQL to optimise posting schedules and media mix.
Built dashboards in Excel, Tableau, and Power BI that grew audience reach by 30%.
I'm after applied research and ML engineering internships — LLMs, agents, evaluation,
or production systems. I'm equally happy training a model and shipping the service around it.
Send me what you're working on.