NOORREHMAN
Experience
Where I've worked
Freelance research, applied ML internships, and production AI systems shipped to real users.
Freelance ML & AI Developer
Independent — Remote · Clients under NDA
- Translated dynamical systems models into code, running simulations and experimental validation for independent research.
- Built an ensemble gold-price forecaster (custom PyTorch transformer, XGBoost, LSTM) with automated BUY/SELL/HOLD signals and Reddit sentiment analysis — full-stack, in-house models, zero external LLM dependency.
Full Stack AI Developer Intern
Noble Platforms — Islamabad · On-site
- Core contributor to The Noble Islam, a live multilingual AI-powered Islamic knowledge app (Quran/Hadith Q&A, English + Urdu) shipped on Google Play.
- Built the AI chatbot using LLM-based retrieval (RAG) over verified religious sources, including multimodal verification of image, video, and audio content.
- Built end-to-end data acquisition and backend systems (Selenium, BeautifulSoup, Playwright, Node.js, Next.js, Python), powering real-time chatbot interactions in production.
Data Science & ML Engineer Intern
Pakistan Bureau of Statistics — Islamabad · On-site
- Automated the national data-cleaning pipeline (imputation, conditional logic, ML-based outlier detection), cutting processing time from hours to under 10 minutes.
- Deployed live Dash and Power BI dashboards with real-time PostgreSQL connectivity, surfacing national KPIs to government decision-makers.
- Presented the system and conducted two workshops on the data-cleaning platform at DataFest 2025.
Projects
Things I've built
From medical ML research to production dashboards — 12 projects across AI, data, and full-stack systems.
ResearchDEXTERA
Skeletal hand-pose HOI recognition across egocentric and exocentric video
TLDR Built an HOI recognition system to classify human-object interactions from raw hand motion. Using the EgoExo4D dataset, I engineered kinematic, spatial, and DFT frequency features, added RS-DACT temporal pooling, and trained class-balanced Random Forests with 5-fold cross-validation across cooking, medical, repair, and music activities.
LiveVoltaIQ
End-to-end consumer outage prediction platform
TLDR Built a full-stack platform so utility customers can see outage risk before it happens. I engineered a pipeline from live weather ingestion through ETL to a gradient-boosted forecasting model, then shipped it as a FastAPI backend with a Next.js dashboard. CV R² = 0.921, MAE = 0.48 hrs.
ResearchOIFA
ODE-informed Parkinson's progression modelling
TLDR Built a framework to predict Parkinson's progression earlier than voice analysis alone allows. I encoded per-patient ODE trajectories as ML features on top of XGBoost, validated with SHAP. Early-window R² = 0.848 vs. voice-only R² = −0.448, and identified 5 distinct PD phenotypes across 42 patients.
LiveGENESIS
AI that makes thinking visible
TLDR Built a tool that shows AI reasoning instead of hiding it. Given a life decision, Genesis generates 6 distinct future paths as an interactive constellation graph, each with editable assumptions, confidence scores, and trade-offs. Built with Next.js, React Flow, Zustand, and an OpenRouter-backed reasoning engine with graceful fallback.
IEEE iCoMET 2026Site EUI Forecasting
Climate-sensitive energy use intensity prediction
TLDR Built a model to forecast building energy use even for buildings with no historical data. I engineered climate-sensitive thermal stress indices and trained separate CatBoost models to solve the cold-start problem directly, rather than defaulting to a shared baseline. Paper accepted at IEEE iCoMET 2026 on the WiDS dataset.
AcademicPIGMENTA
Neural style transfer using only classical image processing
TLDR Built a neural style transfer system under a strict constraint: no PyTorch or TensorFlow allowed. I ran nine pretrained models through OpenCV's cv2.dnn module for inference, then layered LAB colour-preservation, GrabCut segmentation, and Haar-cascade region control entirely with classical image processing on top of that one call.
LiveSentiment-Driven AI Melody Generator
Text-to-music pipeline with real-time emotion mapping
TLDR Built a pipeline that turns written emotion into original music. I detect sentiment with VADER and Bayesian Networks, map it to musical attributes via fuzzy logic, then generate melodies through Markov models synthesized with music21. Shipped as a live Streamlit app — 8–10 sec processing, 90% user satisfaction.
AcademicDeadlock Guard
ML-based deadlock forecasting for OS resource allocation
TLDR Built a system to catch OS deadlocks before they freeze a process, rather than detecting them after the fact. I modeled mutual exclusion, hold-and-wait, no preemption, and circular wait as learned ML features on simulated resource allocation data, moving prediction ahead of static rule-based checks entirely.
Client ProjectMarket Price Analytics System
Client project for Pakistan Bureau of Statistics
TLDR Built a system for PBS to compare official and real market prices at scale. I scraped live product pricing from Metro Online Store using an API with a Selenium fallback for resilience, then designed KPI dashboards and predictive models that fed directly into data-driven policy recommendations.
LiveNational Development Hub
Long-term forecasting aligned with UN SDGs
TLDR Built a forecasting framework projecting Pakistan's crime, migration, and food security trends toward 2030, aligned with UN SDG targets. I delivered it as a live R Shiny dashboard, then paired the technical output with a policy-oriented eBook translating the forecasts into concrete, actionable recommendations for planners.
ResearchCRIMETRIX
OLTP vs. Star vs. Snowflake schema performance study
TLDR Ran a controlled experiment comparing OLTP, star, and snowflake schemas on 8.5M real Chicago crime records in BigQuery. I executed 42 queries measuring slot time, JOINs, and complexity. The star schema hit zero JOINs for 85% of tasks and cut compute time up to 12x versus OLTP.
AcademicSpotify Mood Prediction
Mood prediction and clustering from audio features
TLDR Built a pipeline to organize songs by mood instead of genre. I extracted audio features — tempo, energy, valence — from Spotify's API, then trained Random Forest and SVM classifiers alongside K-Means clustering to power smarter, mood-based playlist recommendations rather than relying on manual tagging.
Skills
Technical expertise
ML & AI
Data Science
Data Engineering
Backend & APIs
Frontend & Visualization
Automation & Workflows
Deployment & Tools
Applied Modeling
Education
Where I studied
BS Data Science
Air University, Islamabad
HSSC, Computer Science
Punjab Group of Colleges (Cantt Campus), Rawalpindi
SSC, Computer Science
Ammar Khan Shaheed Model School For Boys G-10/3, Islamabad
Contact
Let's build something intelligent together.
Open to freelance projects, internships, and full-time/part-time roles in DS, AI, ML, and data engineering.