ML engineer, evaluation researcher
I build systems that test what models actually know, and products that put them to work. Benchmarks, pipelines, and full-stack tools, mostly in Python. I finish things.
Open to ML engineering internships and research collaborations
I'm a Software Engineering student at the University of Sialkot in Pakistan, building ML systems alongside coursework rather than waiting until after it. Most of what I know came from shipping something and then finding out what I got wrong.
My main interest is evaluation: how you tell whether a model is reasoning or pattern matching. That's the thread running through CognitivePrism, which is where most of my time goes right now. The rest of my work is applied, data pipelines, backends, and AI products that need to survive contact with real users.
I care about benchmarks that measure something real, models that do something useful, and code someone else can read six months later.
An adversarial benchmark for testing whether frontier models hold up under epistemic pressure. v2.0 runs 604 items across five tracks, invented phonological grammar, epistemic archaeology, adversarial camouflage, constraint lattices, and institutional cognition, scored by seven functions built for the Google DeepMind × Kaggle benchmarking competition.
I'm currently expanding coverage to 50+ models, with the goal of submitting to the NeurIPS Evaluations & Datasets track in 2027. Open to working with a research lab on the scale-up.
A production chatbot built during my internship at Zidio. Flask REST backend with an external LLM, vanilla JS frontend, conversation context that persists across turns, sentiment analysis on incoming messages, and a Chart.js dashboard for usage analytics. Handles English, Hindi, and Hinglish, which meant the language detection had to tolerate code-switching mid-sentence.
A full-stack ecommerce platform. React frontend on Cloudflare, Python backend, relational schema on Supabase. Currently working through auth and login latency, which is where most of the interesting problems have been.
Classification pipeline predicting churn from customer behaviour data. SQL extraction, pandas cleaning, scikit-learn modelling. Built at Zidio.
LightGBM and XGBoost ensemble for trip duration prediction, with haversine distance features, KMeans geo-clustering, and cyclic time encoding.
Scores B2B leads by conversion likelihood using a self-hosted open-weight LLM, orchestrated in n8n. Owners get a plain-language summary of where each lead stands, not a raw score.
Offline AI coding assistant built on Google Gemini, for restricted or air-gapped development environments where a hosted IDE assistant is not an option.
Real-time augmented reality backend on Supabase, with class-based architecture and live database sync between clients.
Graph-based pathfinding across a multi-floor campus in C++. Dijkstra for shortest routes, BFS for floor search, and fuzzy matching so misspelled room names still resolve.
CRM data pipeline in Python and T-SQL with automated reporting, built and queried through Azure Data Studio.
Discrete 74HC-series logic build with an ESP32 verification layer that checks every input combination against expected output. Documented with full circuit diagrams.
Data Science & Analytics Intern · Bengaluru
Benchmarks and competition pipelines
BS Software Engineering, expected Fall 2028
Coursework in AI and ML systems, data structures and algorithms, digital logic, and software architecture.
AI/ML Professional Certificate
Applied machine learning engineering, model deployment, and Azure AI services.
I'm looking for ML engineering internships, research collaborations, and people working on evaluation. If you're building something worth thinking about, send me a note.