Resume

• Built 5+ production ML models reducing processing time by 40%
• Developed multi-agent systems using LangChain and AutoGen with 92% task completion accuracy
• Implemented automated workflows with Model Context Protocol, saving 15 hours/week

• Quantitative coursework in remote sensing, spatial statistics, and Python-based satellite-data modeling

Self-serve MLOps platform on OpenShift turning an unedited notebook or an externally trained model file into a versioned, evaluated, and monitored deployment

• Built a self-serve platform where an unedited notebook or an externally trained model file becomes a versioned, evaluated and monitored deployment; 23 model versions across 19 models, 9 serving live in production
• Eliminated train/serve version skew by executing uploaded code on the serving image itself, contained in a pod with no service-account token, no secrets, and a 30-minute / 4 Gi fence
• Gated every serving change behind a reviewed pull request carrying the requester's identity (Argo CD GitOps), with a seeded 20% holdout scored per run, drift alerting, and MLflow mirroring within 2 minutes

Technologies: OpenShift, Kubernetes, Argo CD, GitOps, MLflow, Python, Jupyter

Matched-pair reliability study of a spiking neural network and its ReLU twin on DVS128Gesture, with a live demo that runs both models on damaged event-camera recordings

• Held architecture, parameter count, data, seed, loss, and schedule constant and swapped only the neuron model, then measured accuracy, calibration, robustness under four corruptions, temporal sensitivity, and accounting-model energy, with bootstrap 95% CIs on every number (n=288)
• Found the spiking model gives up 3.5 accuracy points (McNemar p = 0.021) but ties on calibration (ECE 0.040 vs 0.039), leads by up to 28 points under sensor noise with confidence that tracks its accuracy while the ANN stays up to 34 points overconfident, and costs 18.7x less on the standard energy accounting
• Shipped a live demo on Modal (CPU, memory snapshot, scale to zero) behind a keyed proxy in the portfolio, replaying both models frame by frame on 33 held-out recordings with the published curves alongside

Technologies: Python, PyTorch, SpikingJelly, NumPy, Modal, FastAPI, Next.js

QLoRA fine-tuning of Qwen 2.5 7B Instruct for natural Urdu, Roman Urdu, and Urdu/English code-mixed chat, with RAG-aware retrieval over Urdu Wikipedia

• Fine-tuned Qwen 2.5 7B with QLoRA (154 MB adapter, 0.82% of parameters, 8.56 GB peak VRAM) across three versions, built end-to-end for under $60
• Achieved 79.5% pairwise win rate over the base model across two independent LLM judges on a 100-prompt evaluation set, recovering every regression from the prior version
• Ran a data-centric iteration loop that lifted code-mixed task win rate from 0% to 80% and made retrieval deployable, cutting the base model's Chinese-fallback on Urdu context from 45/100 to 0

Technologies: Python, PyTorch, Unsloth, QLoRA / PEFT, Hugging Face, Modal, Gradio, Haystack, Qdrant

ML system predicting PERM labor certification outcomes with SHAP explainability

• Built threshold-tuned Gradient Boosting classifier achieving 73.2% accuracy with 61% denied recall on 25K PERM records
• Implemented SHAP TreeExplainer with feature mapping to provide per-prediction explanations across 10 input features
• Designed 5-stage modular MLOps pipeline (ingestion, validation, transformation, training, evaluation) with model promotion gating

Technologies: Python, Scikit-learn, XGBoost, LightGBM, CatBoost, SHAP, FastAPI, Docker

Deep learning system for facial expression recognition in group photos using ensemble models

• Built emotion classification model achieving 80% accuracy on RAF-DB dataset across 7 emotion classes
• Implemented ensemble learning combining ResNet-18 and EfficientNet-B2 to handle severe class imbalance
• Deployed Flask web application with MTCNN face detection on Hugging Face Spaces using Docker

Technologies: PyTorch, Flask, MTCNN, ResNet-18, OpenCV, Docker, Hugging Face

ML regression model for agricultural yield forecasting using satellite imagery and climate data

• Built supervised ML model achieving 0.137 t/ha prediction error on test set
• Engineered features from multi-spectral satellite imagery and climate variables using Google Earth Engine
• Tuned hyperparameters under cross-validation

Technologies: Scikit-learn, Python, NumPy, Pandas, Google Earth Engine, Feature Engineering

Teacher rating and review platform for educational institutions

• Built full-stack web application with REST APIs for real-time data synchronization
• Achieved 60% bundle size reduction through code splitting and lazy loading optimizations

Technologies: React, TypeScript, Supabase, TanStack Query, Tailwind CSS, Vite

Geospatial AI system analyzing 22 years of satellite data for groundwater prediction (2002-2024)

• Developed ML regression models achieving R²=0.89 for groundwater level predictions across 145 districts
• Deployed Flask REST API serving ML models with real-time GRACE satellite data processing
• Engineered feature extraction pipeline processing 22 years of geospatial time-series data

Technologies: Scikit-learn, Flask, Google Earth Engine, React, GRACE/GLDAS

ML-driven spatial optimization for Electric Vehicle infrastructure planning

• Implemented weighted scoring algorithm processing demographic, economic, and infrastructure layers for 5 tehsils
• Ranked candidate sites with multi-criteria decision analysis to a 90%+ coverage target

Technologies: Python, Scikit-learn, QGIS, ArcGIS, OpenStreetMap, Multi-criteria Analysis