ML Engineer at DeepEdge. Leading anti-UAV tracking systems deployed at 3 defense installations. Real-time inference on NVIDIA Jetson edge devices. Sub-pixel object detection < 10ms latency.
Contributor to PyTorch Lightning (27K★), Hugging Face (135K★), KerasCV / KerasNLP. Augmentation layers used by 35K+ engineers. 1M+ monthly developer reach.
Detection range extension via sub-pixel multi-algorithm ensemble.
Per-frame inference on NVIDIA Jetson. Real-time 30fps edge.
Leading end-to-end development of production anti-UAV tracking systems for defense installations. Real-time computer vision on edge. Full ownership from research to deployment.
Built production ML systems for enterprise clients — recommendation engines, predictive models, and multilingual RAG pipelines. Shipped measurable business impact at scale.
Built cross-framework ML infrastructure for the Ivy compiler. Shipped TensorRT converters and rigorous testing for production-grade code used by thousands of engineers.
Python, C++, TypeScript, JavaScript, SQL, Bash.
PyTorch, PyTorch Lightning, TensorFlow, JAX, Scikit-learn, XGBoost, LightGBM.
YOLO (v5/v7/v8), Detectron2, OpenCV, Kalman filters, TAPNet, CoTracker, thermal imaging, sub-pixel tracking.
TensorRT, CUDA kernel fusion, quantization, pruning, NVIDIA Jetson, <10ms real-time inference.
Hugging Face Transformers, PEFT, LangChain, LangGraph, RAG, RLHF (PPO / DPO), vector databases.
Docker, Kubernetes, FastAPI, CI/CD for ML, AWS, PostgreSQL, Redis, FAISS.
From-scratch RLHF pipelines using PPO and DPO for LLM alignment. Comparative analysis of reward-model-based PPO vs Direct Preference Optimization — trade-offs in stability, compute efficiency, and output quality.
Lightweight vector DB built from scratch using Hierarchical Navigable Small World (HNSW) graphs. No LangChain, no Pinecone — just the algorithm and NumPy. Minimal dependencies, easy integration.
From-scratch implementations of foundational deep learning papers in PyTorch. Not API wrappers — actual algorithm reimplementations to understand the math underneath.
Natural language analytics over structured flight data. RAG architecture with DeepSeek-R1-Distill-LLaMA-70B and LLaMA 3.2:3B. ChromaDB retrieval, FastAPI serving.
Fine-tuned ALBERT on 500K Amazon listings across 27 categories with 1:150 class imbalance. 89% F1 — beat BERT by 12% and production baseline by 10%. TensorRT deployment.
Gated RNNs analyzing children's physical activity patterns for early detection of problematic internet use. Healthcare ML with real clinical relevance.
Top 10 teams worldwide. Built disaster response system using satellite imagery + EfficientNet — 91% IoU.
Top 1% ranking among 2,500 participants in a machine learning competition.
Top 1.2% in pharma/healthcare data science competition.
Top 500 women technologists globally. 0.5% acceptance rate.
Consistent problem-solving practice. Data structures, algorithms, systems.
Mentored 15+ contributors on OpenSSF project. 25+ merged PRs during the event.