retrieval · learning · deployment
machine learning engineer, karachi
© 2026 Shaikh Muhammad Zain
ResumeApproach
0.24 → 0.74.
Precision@50 at FlyRank, over the hand-written baseline
Every project starts with the data: what it holds, what it can't answer, and a baseline to beat.
Then the model, then an honest check. When a result doesn't survive the check, I say so and write down why.
Then it ships: an API, a demo you can click, and a README with the numbers.
Selected work
1 / 6
01
Student performance intelligence platform · 2026
Flags students likely to average below 60 before any exam is taken, explains every prediction with SHAP and audits the model for fairness across student groups. Ten model families are benchmarked from one training command.
02
Search ranking model and content archetype study · 2026
Built classification models on real anonymised search data and scaled the pipeline from a 30k-row starter set to a 79M-row warehouse. The capstone clustered 18,752 pages into behavioural archetypes; five survived an ablation test and the sixth was withdrawn.
03
Hybrid retrieval and RAG with citation checks · 2026
Searches 2,500 commit messages from huggingface/datasets by meaning, combining BM25, LSA and sentence-transformer embeddings in FAISS. A language model answers with commit citations, and every cited id is checked against what was retrieved.
04
Rule-based chatbot, DecodeLabs internship · 2026
Answers with rules instead of a model, so every reply traces back to a line in intents.json. Input is matched in five tiers, from exact lookup to TF-IDF similarity, and the web interface shows which tier answered.
05
KNN with an explainable web app, DecodeLabs internship · 2026
A K-nearest-neighbours model behind a JSON API and a page that shows the neighbours that cast each vote. Every chart is drawn in the browser from the live model.
06
Final year project, NED University · 2026–27
Middleware that wraps LangGraph agents, flags hallucinations at each step using log-probability screening and a fine-tuned ModernBERT classifier, and tries to repair them without changing the agent.
Python first. Most projects run from a notebook to an API in the same repository, with tests and CI.
Retrieval: BM25, dense embeddings and FAISS, each benchmarked on the same questions before I pick one.
Models: scikit-learn and XGBoost for tabular data, PyTorch and Transformers for text, SHAP when someone needs to know why.
Shipping: FastAPI, Flask and Streamlit, deployed on Vercel and Render.
Measured, then shipped.
Final-year Computer Science student at NED University in Karachi, specialising in artificial intelligence and graduating in August 2027.
I work on retrieval systems and supervised learning, and on getting models out of notebooks into something a person can run. Right now I'm building StepGuard, my final year project.
Coursework: Machine learning · Data mining · Parallel and distributed computing · Algorithms and data structures · Computer networks.
FlyRank, machine learning intern
DecodeLabs, artificial intelligence intern
Karachi Development Authority, information technology intern
NED University of Engineering and Technology
DJ Sindh Government Science College
Al-Jauhar Grammar School
Google AI Professional Certificate (Coursera) · Machine Learning Certification (FlyRank) · AI Fluency Certification (FlyRank) · Virtual Internship, Artificial Intelligence (DecodeLabs) · AI for Research and Insights · AI for Brainstorming and Planning · AI for Content Creation · Web Development Course (Aptech Computer Education).
“I ship models past the notebook stage.”
Remote or international ML internships and junior ML engineer roles. Email is the fastest way to reach me.
Back to top