Full-stack software engineer building web applications, backend systems, and data-driven products.
- 💻 Full-stack and backend-focused developer
- 🚀 Working with TypeScript, Next.js, React, Node.js, and PostgreSQL
- 🔐 Experience with authentication, authorization, RBAC, and OAuth/SSO
- 📊 Built analytics platforms and backend services using ClickHouse and Hono
- 🤖 Previously worked on an AI-based auto-parallelization project at C-DAC using DeBERTa-v3 and PyTorch
Languages
TypeScript · JavaScript · Python · SQL · C++
Frontend
React · Next.js · Tailwind CSS
Backend & Data
Node.js · Hono · Express · PostgreSQL · ClickHouse · MongoDB · Prisma · Convex
Other
Git · Postman · Vercel · PyTorch
A real-time collaborative code editor built with React, Node.js, Socket.io, and Yjs CRDTs.
- Real-time collaborative editing with conflict-free synchronization
- Live cursors, presence, and in-room chat
- JWT authentication for REST and WebSocket connections
- Persistent sessions with MongoDB
- Sandboxed live preview for HTML, CSS, and JavaScript
AutoPragma — AI-Based Auto-Parallelization · Model
During my internship at C-DAC, I fine-tuned a DeBERTa-v3 model on 42,360 labeled C/C++ loops to classify OpenMP parallelizability, and generated OpenMP pragma recommendations from its predictions.
- 81% accuracy and 0.77 F1 on a 5,296-loop held-out test set
- Evaluated across NAS Parallel Benchmarks, PolyBench, and SPEC OMP
- Degrades under distribution shift: on SPEC OMP it scored 82.5% accuracy but fell below the 86.4% majority-class baseline — so every suite is reported against that baseline rather than the favorable benchmark
A leak-free scikit-learn regression pipeline over the Ames Housing dataset (1,460 homes, 79 features), wrapped in a deployed Streamlit app.
- Imputation, scaling, and encoding are fit inside each CV fold, never on the full dataset
- Every error metric reported in dollars — R² 0.914, MAE $15,407 on the held-out test set
- Model selection ignores R² gaps smaller than the fold-to-fold noise and breaks ties on MAE, which picked the model that turned out better on unseen data
- pytest suite and CI on every push
- 💼 LinkedIn: linkedin.com/in/harsh-agrawal-dev
- 🌐 Portfolio: https://agrawalharsh.in/


