Portfolio
Projects
Systems I've built — each with source code and detailed documentation.
Scalixor — Distributed Real-Time Communication Platform
Horizontally scalable, fault-tolerant platform for group chat, 1:1 WebRTC video calling, and collaborative code editing across multiple backend instances.
- ◆Architected a horizontally scalable, fault-tolerant, real-time platform supporting group chat and 1:1 WebRTC video calling across multiple backend instances.
- ◆Implemented Redis Pub/Sub for cross-server message broadcasting, signaling relay, and global online coordination.
- ◆Extended the system with a multi-room collaborative code editor (Monaco) synchronized via Redis Pub/Sub.
HealthVault — Secure Healthcare Data Sharing Platform
Privacy-preserving medical records platform with AES-256-GCM envelope encryption, patient-controlled consent, RBAC, and blockchain audit trail.
- ◆Built a privacy-preserving medical records platform with AES-256-GCM envelope encryption (per-record DEK + master KEK), patient-controlled consent management, and RBAC across Patient, Doctor, and Admin roles.
- ◆Implemented TOTP-based multi-factor authentication, JWT access/refresh token flow, and Attribute-Based Encryption (ABE) policies for fine-grained record-level access control.
- ◆Designed a SHA-256 hash-chain blockchain audit trail for tamper-proof logs with admin-facing integrity verification.
RuralConnect — Full-Stack Rural Community Platform
Full-stack platform for rural communities featuring AI chatbot, real-time chat, video calling, payment integration, and blog CMS.
- ◆Built a full-stack platform for rural communities with an AI chatbot (Gemini API), real-time chat via Socket.io, and peer-to-peer video calling using WebRTC.
- ◆Implemented three-tier JWT auth (user, expert, admin), Razorpay payment integration, and a blog CMS with rich-text editing and S3-backed media uploads.
- ◆Designed a scalable backend with Prisma ORM, Redis caching, and containerized the full stack using Docker Compose with PostgreSQL and Redis services.
Personal Finance AI Agent
Multi-tenant financial platform with workspace isolation, RBAC, Redis caching, BullMQ workers, and Python FastAPI AI orchestration.
- ◆Architected a multi-tenant financial platform with workspace isolation, role-based access control, and stateless REST APIs secured using JWT authentication.
- ◆Implemented Redis caching and BullMQ background workers to support asynchronous AI tasks and recurring jobs.
- ◆Integrated a Python FastAPI microservice for AI agent orchestration and containerized the system with Docker.
Distributed File Repository System
Secure multi-client file server using raw TCP sockets with SSL/TLS encryption and concurrent thread-per-connection handling.
- ◆Built a secure multi-client file server using raw TCP sockets with SSL/TLS encryption and concurrent thread-per-connection handling.
- ◆Implemented store-and-forward messaging with mutex-based synchronization to prevent protocol-level race conditions.
Paradigm Shift Ranker — Temporal Graph Information Bottleneck
Temporal GNN framework detecting paradigm-shifting research papers via latent-space displacement — published at ACM GRADES-NDA 2026.
- ◆Designed a Temporal GNN framework (GraphSAGE encoder + Variational Information Bottleneck) achieving AUC 0.9989 and perfect NDCG@20 on OGB-Arxiv.
- ◆Introduced Innovation Velocity — an L2-norm displacement metric in the IB-compressed latent space — surfacing high-impact papers 2–3 years before conventional methods.
- ◆Deployed the system end-to-end with FastAPI, Redis caching, Kubernetes auto-scaling, and GitHub Actions CI/CD; published as an ACM GRADES-NDA 2026 workshop paper.
Automatic Question Generation Without LLMs
Rule-based NLP pipeline using spaCy to auto-generate educational questions with Bloom's Taxonomy-aligned difficulty classification.
- ◆Built a rule-based NLP pipeline using spaCy (POS tagging, NER, dependency parsing) to auto-generate educational questions from raw text — no LLMs or neural networks used.
- ◆Engineered a multi-feature difficulty classifier mapping questions to Bloom's Taxonomy cognitive levels.
- ◆Evaluated against SQuAD 2.0 (130K+ QA pairs) using BLEU and ROUGE metrics; served via Flask REST API with Next.js frontend.