Do human reviewers catch model errors?
When people approve AI judgments at high rates, it is unclear whether the model is right or the reviewer is deferring. I want to measure the difference.
Evidence: GradexAI human-review loop, 1,946 gradings
Open to AI safety research fellowships and US faculty roles
Gul Sher Baloch
Applied ML researcher · Associate Professor, Sukkur IBA University · Co-founder & CEO, Leaveit2AI LLC
My PhD was about recovering clean signals from noisy ones. For the last two years I have run LLM systems for teachers and doctors, where a wrong answer costs a real person. The question I keep returning to: how do humans tell when the model is wrong?
Research focus
Oversight and reliability of LLM systems
Three threads connect my earlier signal-processing work to the AI safety questions I want to study next.
Where should an LLM be allowed to answer?
In medical exam prep I kept the model behind the scenes, checking and tagging content, rather than answering doctors directly.
Evidence: GulfMedExams pipeline and error resolver
Learning when the labels are wrong
Noisy labels in vision data and noisy preferences in human feedback are the same problem at different scales.
Evidence: label-noise framework, IEEE Access 2025
Selected work
What I built, and what it showed
Numbers below come from each product's own dashboards.
GradexAI
PausedLLM exam creation and grading, with a teacher reviewing every grade
- Built
- LLM-generated rubrics, OCR for handwritten scripts, and outcome analytics for accreditation reports.
- Reach
- 17 institutes, 1,516 users, 270 exams, 1,464 student submissions.
- Lesson
- Institutions would not change paper-exam habits; demo enthusiasm is not adoption. Validate with real users at every iteration.
Approved 91.7%Not approved 8.3%
Open question: how many of the approvals were wrong? That is the oversight gap I want to measure.
GulfMedExams
LiveGulf medical licensing exam preparation for doctors
- Built
- LLMs restructure, quality-scan and tag recalled questions; two-stage deduplication (OpenAI models plus embeddings) builds a high-yield "Repeat Vault".
- Decision
- No AI-generated questions and no user-facing chat. One hallucination would cost credibility with doctors.
- Result
- $12K+ revenue in under 4 months; 265 paid users in 40 countries; 39 of 40 surveyed users reported passing.
Fixed 189Dismissed as correct 105Not auto-closed 22
Pipeline scale: 6,946 doctor-verified MCQs, tagged by topic and subspecialty.
AI Maths Tutor
ShelvedO-Level maths prototype
- Built
- Worked solutions loaded and checked in advance; a student clicks the step they are stuck on and a chatbot explains it.
- Why
- Verify first, then let the model explain. Nothing a student sees is generated live.
Agronomist
First prize, DICEStudent AI flower-identification app I product-led as faculty mentor
- Data plan
- Images on different phones and resolutions, from different angles, at different times of day and in different weather, so the model holds up outside clean photos.
- Result
- DICE Innovation First Prize and a place in the university's startup incubator.
A project I would start with
Do overseers catch wrong LLM grades?
When a capable model grades work and a human approves it, how often is the approval a rubber stamp? GradexAI gave me the setting; this project would give it a measurement.
- Plant known errorsSeed LLM-graded short answers with errors of graded difficulty, using public or synthetic student work.
- Measure catch ratesCompare human reviewers with weaker models acting as reviewers, across error types.
- Test interventionsShowing the rubric, a model-written critique, or two models debating the grade.
- Release itPublish the dataset, evaluation code and results openly.
Publications
Selected papers
Full list and citation counts on Google Scholar. Code links go to the original research repositories.
- 2025Robust learning
A robust deep learning framework for mitigating label noise with dual selective attention
Hyder, Baloch, Batool, Kim & Byun · IEEE Access
- 2022Code
Neuromarketing solutions based on EEG signal analysis using machine learning
Ullah, Baloch, et al. · IJACSA 13(1)
- 2021Paper repo
Classification of alcoholic EEG signals using wavelet scattering transform-based features
Buriro, Ahmed, Baloch, et al. · Computers in Biology and Medicine 139
- 2021Code
Fungal blast disease detection in rice seed using machine learning
Kumar, Baloch, et al. · IJACSA 12(2)
- 2021Code
Analysis of text feature extractors using deep learning on fake news
Ahmed et al. · Engineering, Technology & Applied Science Research 11(2)
- 2019Medical imaging
Impact of image enhancement technique on CNN model for retinal blood vessels segmentation
Soomro, Baloch, et al. · IEEE Access 7
- 2018Code
Residual correlation regularization based image denoising
Baloch, Ozkaramanli & Yu · IEEE Signal Processing Letters 25(2)
- 2017Image processing
Image denoising via correlation-based sparse representation
Baloch & Ozkaramanli · Signal, Image and Video Processing 11(8)
- 2017Conference
Coupled K-SVD dictionary learning algorithm in wavelet domain for single image super-resolution
Ahmed, Baloch & Ozkaramanli · IEEE International Symposium on Imaging Systems and Techniques
Experience
Research, teaching and building
Education
- PhD, Image Processing, Eastern Mediterranean University, Cyprus (2018)
- ME, Communication Systems & Networks, Mehran UET
- BE, Telecommunication Engineering, Mehran UET
Recognition
- IEEE Senior Member
- MIT CSAIL Professional Certification, ML & AI (2019)
- Research Achievement with Distinction, EMU (2018)
- DICE Innovation First Prize
- Invited speaker, PMI Michigan & PMI Buffalo (2025)
-
2024 – now
Co-founder & CEO
Leaveit2AI LLC (USA)
- Lead a 4-person team that tested five AI products, bootstrapped for two years; now funded by GulfMedExams revenue.
- Design the LLM pipelines, evaluation and human-review steps behind each product.
-
2011 – now
Associate Professor & Program Coordinator
Sukkur IBA University, Pakistan · Lecturer 2011, Assistant Professor 2018, Associate Professor 2021
- Teach AI, Machine Learning, Deep Learning, Python and Data Science, with labs on LLM integration, RAG and AI agents; 90%+ student evaluations in AI courses for three consecutive years.
- Supervised 30+ applied AI/ML projects; PI on HEC and NGIRI research grants; HEC-approved PhD supervisor.
- Reviewer for IEEE Transactions on Image Processing and IEEE Access; editor, ETASR.
- Lead an accredited engineering program; built a paperless course-file system supporting PEC's highest-level OBE accreditation.
-
2019 – 2021
Founder & Lead Coordinator, AI for Schools
With ReadyAI (USA)
- Ran what ReadyAI called Pakistan's first K-12 AI teacher training; WAICY Country Organizer for Pakistan.
-
2014 – 2018
Research & Teaching Assistant
Eastern Mediterranean University, Cyprus
- PhD research on sparse-representation and correlation-based image denoising and super-resolution.
Tools
What I work with
Programming
Python (NumPy, Pandas, scikit-learn, OpenCV, Matplotlib), MATLAB, SQL, Git, REST APIs, MongoDB
ML and LLMs
Deep learning (TensorFlow, Keras), Hugging Face, OpenAI API, embeddings, RAG, LangChain, AI agents, OCR
Research practice
Experiment design, model evaluation, human-in-the-loop evaluation, dataset design, peer review, grant writing
Contact
Get in touch
gulsher9931@gmail.com
Based in Pakistan, relocating to Houston, TX. US work authorization: EB-2 NIW green card, immigrant visa interview completed, issuance pending. No employer sponsorship required.