· Building distributed systems, AI infrastructure, and developer tools

Hey Folks! I'm a 3rd-year mechanical engineering student at MNNIT Allahabad. I build and tune deep learning models and ship them end-to-end: training pipelines, FastAPI services, Docker, and cloud deployments.

My work sits at the intersection of computer vision, ML systems, and production deployment. I care about building things that actually run in production — not just notebook demos.

• Open to collaboration opportunities in distributed systems, AI infrastructure, and full-stack engineering

• YOLOv8-nano + ByteTrack pipeline — real-time vehicle detection with persistent multi-objecttracking IDs maintained across 200+ frames.
• Dual-line crossing counter with per-vehicle dedup via Python set operations — 96.2% accuracy, 100% recall validated on 2,208-frame manually-annotated ground truth.
• End-to-end deployment — raw video → annotated MP4 + per-event CSV log, served via Gradio onHuggingFace Spaces with matplotlib analytics dashboard.

Python PyTorch YOLOv8 ByteTrack OpenCV Gradio HuggingFace Spaces matplotlib

• ResNet18 transfer learning — fine-tuned ImageNet-pretrained ResNet18 on FER2013 across 7 emotion classes; 68.7% test accuracy validated on 7,178-image held-out split (beats published baseline of 60–65%).

• GradCAM interpretability layer — generates heatmap overlays showing which face regions (eyes,mouth, eyebrows) drove each prediction; integrated as a toggle in the live demo.

• End-to-end Gradio app — webcam capture + image upload + per-class probability breakdown, deployed on HuggingFace Spaces (free CPU tier) with git-lfs for the 43MB checkpoint.

  Python PyTorch ResNet18 GradCAM Gradio HuggingFace Spaces MPS

• 5-service production stack — React frontend on Vercel + 4 containerised FastAPI services on Render: JWT-authenticated gateway with RBAC, PyTorch encoder (ResNet50 / ViT), FAISS retrieval engine, and evaluation service.
• Cross-modal retrieval pipeline — input tile → 512-D embedding → FAISS IndexFlatIP cosine search with sub-100ms latency, returning ranked matches with per-result explainability (NDVI, geometric alignment, embedding distance).
• Vercel rewrite proxy routing /auth, /api, and /health from the SPA origin to the backend — eliminated CORS preflight overhead and unified the browser-facing domain.

Python PyTorch FAISS FastAPI React Docker Vercel Render

Languages       C  C++  Python  JavaScript  TypeScript  HTML/CSS
ML / AI         PyTorch  Scikit-Learn  YOLOv8 (Ultralytics)  ByteTrack  OpenCV  NumPy  Pandas
Deployment      Hugging Face Spaces  Gradio  matplotlib  Git  GitHub  Vercel Render
Learning        NLP  Transformers (BERT/DistilBERT)  Embeddings  RAG  LangChain
Interests       Computer Vision  Deep Learning  NLP  Information Retrieval  MLOps  DSA
CP              400+ problems solved on LeetCode & CodeChef

_{📫 jayaarora2402@gmail.com · linkedin.com/in/jaya-arora-6892a93a0 · github.com/Jaya242}