Rusham ElahiMachine Learning Engineer

Ranked underserved districts across South Asia by combining travel time to facilities, gridded population, and nighttime lights into a single accessibility score.

• • Built the Google Earth Engine implementation and the scoring method behind the study, combining travel time, gridded population, and nightlights to rank underserved districts across South Asia
• • Led data preprocessing, feature engineering, and the geospatial analysis underpinning the published results

Python, Google Earth Engine, QGIS

Geospatial pipeline mapping flood inundation across Punjab and Sindh and quantifying damage to schools and hospitals in flood zones for disaster response.

• • Mapped flood inundation over ~240,000 km² of Sindh and South Punjab using Sentinel-1 SAR imagery in Google Earth Engine; applied speckle filtering, pre/post-flood differencing, and permanent water body removal to generate flood masks
• • Trained YOLOv5 on satellite imagery to detect schools and hospitals in flood zones; performed spatial intersection analysis in QGIS to quantify infrastructure damage
• • Led data preprocessing, feature engineering (including NDWI computation), and geospatial analysis pipelines for disaster response across two provinces

Google Earth Engine, QGIS, YOLOv5, Sentinel-1 SAR, Python

Built a multi-source geospatial ML pipeline to predict business location suitability using satellite imagery and POI data.

• • Collected 12,000+ business POIs via Google Places API; extracted geospatial features (nighttime light, population density, NDWI, land-use) from Sentinel-1, Landsat, and Google Earth Engine
• • Fused multi-source satellite and ground-truth data into a unified dataset for training
• • Trained XGBoost classifier achieving 80%+ accuracy on business location suitability prediction

XGBoost, Google Earth Engine, Python, QGIS, PostgreSQL

Ongoing work applying graph neural networks to business site selection. Represents urban regions as graphs — nodes are places, edges capture geographic and functional proximity — and learns region embeddings that flag high-potential business zones.

• • Modeled urban regions as graphs with nodes representing places and edges capturing geographic and functional proximity
• • Trained a graph neural network to learn region embeddings over real city POI data
• • Compared against clustering and conventional spatial-analysis approaches; evaluation in progress

PyTorch Geometric, Graph Neural Networks, Python, QGIS

Reproduced BIT on LEVIR-CD at F1 0.903 (0.893 reported), then found frozen DINOv2 features transfer better zero-shot than fine-tuned ones.

• • Reimplemented BIT (Bitemporal Image Transformer) in PyTorch and reproduced it on LEVIR-CD
• • Swapped its ResNet encoder for frozen and fine-tuned DINOv2 and DOFA backbones behind a patch-token adapter, to test how geospatial foundation-model features transfer to bitemporal change detection
• • Zero-shot LEVIR→WHU-CD, frozen DINOv2 features held F1 0.781 against 0.704 for the ResNet baseline, while fine-tuning the same backbone collapsed to 0.386, despite the frozen model scoring lower in-domain

PyTorch, BIT, DINOv2, DOFA, LEVIR-CD, WHU-CD

Prithvi-EO-2.0 reached water-class IoU 0.817 against 0.672 for an NDWI baseline on Sentinel-2 flood segmentation.

• • Deployed and independently evaluated NASA-IBM’s Prithvi-EO-2.0 (300M-param ViT geospatial foundation model) for flood-water segmentation on Sentinel-2 imagery
• • Benchmarked the Sen1Floods11 checkpoint against an NDWI water-index baseline on the 89-tile validation split
• • Grad-CAM over the UperNet decoder to surface failure modes; deployed as a Gradio app on Hugging Face Spaces

Python, Prithvi-EO-2.0, TerraTorch, Grad-CAM, Gradio, Hugging Face

Random Forest implemented from scratch in NumPy reached 78.46% accuracy on held-out years 2020–2023.

• • Decision Tree, Random Forest, and Gradient Boosting classifiers implemented from scratch in NumPy
• • Predicts WHO Universal Health Coverage tier (SDG 3.8) from workforce and infrastructure features

NumPy, scikit-learn, pandas

Measures the real-world width and height (mm) of a notebook from a single phone photo.

• • iPhone 12 camera calibration at 0.085 px reprojection error
• • Mask R-CNN segmentation of the notebook (val IoU 0.96)
• • Homography on a credit card reference for mm scale

Python, Mask R-CNN, OpenCV, Camera Calibration, Homography

Pakistani news voice agent: you speak, it fetches Dawn headlines and reads them back over a live voice session.

• • Next.js frontend + Python agent on the LiveKit Agents framework
• • Groq for STT (whisper-large-v3), LLM (llama-3.3-70b), and TTS in one provider
• • Tool-calling for headlines / detail / category filter, with custom rate-limit and interruption rails to keep playback clean

Next.js, Python, LiveKit Agents, Groq (Whisper, Llama 3.3 70B, TTS)

Scrapes medicine prices from online pharmacies and cross-references them against DRAP-regulated prices.

• • Flags pricing anomalies and counterfeit medicines

scikit-learn, Pandas, Streamlit, Gemini API, BeautifulSoup

Building data pipelines and running graph-neural-network inference over multi-million-record scientific datasets.

• • Ran batched graph-neural-network inference to embed ~160k structures for downstream property prediction
• • Built resumable pipelines that pull multi-million-record scientific datasets through public APIs into sharded parquet

Shipping features on a production insurance CRM across Laravel/Angular and React/FastAPI services.

• • Ship features on a production insurance CRM built with Laravel/Angular and React/FastAPI, containerized with Docker on AWS
• • Built AI chatbots and a bilingual client portal for client-facing workflows
• • Delivered an 8-workflow n8n lead-intake pipeline backed by Postgres

Built and deployed custom Whop applications with integrated authentication and payment systems.

• • Built and deployed 7+ custom Whop applications with multi-tenant role-based access control, JWT authentication, and integrated payments
• • Built a generative AI pipeline on the Gemini API with an SSE notification system that pushes real-time progress during long-running batch jobs

Mapped flood inundation across ~240,000 km² of Sindh and South Punjab from Sentinel-1 SAR, and quantified damage to schools and hospitals in the affected zones.

• • Mapped flood inundation from Sentinel-1 SAR in Google Earth Engine using speckle filtering, pre/post-flood differencing, and permanent-water removal
• • Trained YOLOv5 on satellite imagery to detect schools and hospitals in flood zones, then quantified infrastructure damage by spatial intersection in QGIS
• • Built the Earth Engine implementation and scoring method behind the ICONIP 2024 healthcare accessibility study

Built neural networks in TensorFlow with the AI team

Final year project: LocatePro — multi-source geospatial ML for business site suitability.

• • Dean's Honor List — academic excellence, Feb 2022
• • Best Student, CAIR, SCIT — ML & Data Science Innovations, May 2024
• • Most Popular Project Team, SCIT, BNU — ML & Data Science Innovations, May 2024