Hi, I'm Hunain

An MCP server that lets an agent design neural nets on a shared canvas

An MCP server exposing tools that an AI agent uses to design PyTorch architectures, while a human edits the same network live on a React canvas. Shape errors are caught by actually executing the generated model against a dummy batch in a sandboxed subprocess, with a timeout and an address-space cap, returning structured per-node errors instead of raw tracebacks. Emitted code is self-contained as it imports only torch.

Hybrid-retrieval RAG over university handbook

Question answering over NUST undergraduate handbook. The approach of using MinHash and SimHash were a course requirement, we were not allowed to use vector embedding models. The retrieval stack combines MinHash+LSH candidate filtering, IDF-weighted SimHash re-ranking, a TF-IDF baseline and RRF fusion, on top of a MapReduce preprocessing pipeline.

Transformer fusion for a language most models ignore

Progressive DL pipeline for hate speech detection in Roman Urdu, a language 170M+ people speak and is very underrepresented. TF-IDF baselines, then custom neural architectures including CNNs and GRUs, then multilingual transformer fusion. Handles 5+ spelling variants per word and the 63.5% of hate that carries no keyword at all.

Full-stack crime analytics with AI predictions and safe routing

React and Express app with crime data in PostgreSQL, community posts in MongoDB and AI predictions cached in Redis. Leaflet maps, Recharts dashboards, Gemini-powered crime prediction, and a safePath algorithm that routes people through the lowest-crime corridors.

Cron-driven stock prediction pipeline, end to end

OHLCV ingestion from several APIs, preprocessing, Random Forest ticker filtering, then ANN price prediction. Built during the Tensor Labs internship and handled only the automation pipeline part, RF and ANN were provided by client.

Quantifying solar power lost to pollution

ML model combining weather, pollution and irradiance data to predict solar plant output and put a number on the power lost during Pakistan's smog season.

Interactive analytics across a century of Olympic data

Streamlit app visualising athlete demographics, country participation, medal breakdowns and coach analytics across several Olympic datasets.

A banking database built from raw data structures

JavaFX DSA course project built by a team. No external database was allowed, so the banking system was simulated with nested HashMaps and CSV files, including user and admin dashboards and heuristic fraud detection.