Full-Stack & Backend Engineer. Cloud-native backends, edge-native AI, and the open source plumbing in between.
I build full-stack applications where the backend is the hard part, and I like the hard part.
Right now that means a Google Summer of Code 2026 project at OWASP OpenCRE: a multi-stage ingestion filter that decides what security content is worth harvesting, from regex gates through an LLM relevance classifier, with an evaluation harness to prove the recall. Before that, a privacy-first AI inbox that runs inference entirely on Cloudflare's edge, and a Node backend deployed on ECS Fargate with Terraform.
What I care about:
- Architectures that stay simple as they grow
- Backends measured by latency and reliability, not feature count
- AI that respects the data boundary: on-device, at the edge, or not at all
- Open source, and writing down what I learn so the next person goes faster
I contribute where the plumbing is: data pipelines, UI migrations, CI, and docs.
| Org | What I work on |
|---|---|
| OWASP OpenCRE | GSoC 2026 Module B: harvest input and knowledge-queue models, regex and sanitize filter stages, LLM relevance classifier, recall-first evaluation harness |
| Palisadoes Foundation | Talawa Admin: table loader migration, unreachable-code and UI fixes |
| Kestra | Migrating Vue UI components to TypeScript |
| Fastify | Core plugin docs and CITGM coverage for @fastify/sse |
VaultMail · Cloudflare Workers Workers AI Llama 3.1 Durable Objects R2 KV React Tailwind TypeScript
Privacy-first AI email triage that never sends a byte of your inbox to a third-party AI vendor. Priority, category, summary and suggested action are inferred at the edge, summaries are AES-GCM-256 encrypted before storage, and every AI call lands in an auditable gateway log. Live
Theta E-Learning Platform · Node.js Express MongoDB AWS ECS Fargate S3 Terraform GitHub Actions React Vite Playwright
Course platform with lecture video delivery, JWT auth, Razorpay payments and progress tracking. The API runs as a Fargate task behind an ALB in a custom two-AZ VPC, with all infrastructure in Terraform and secrets injected from Secrets Manager at launch. Live app · Frontend repo
Hybrid Temporal Forecaster · Python PyTorch ARIMA CNN-BiLSTM Attention
Three-phase study forecasting hourly electricity demand on 145k rows of PJM East load data. An ARIMA baseline, then a CNN-BiLSTM with additive attention cutting MAE by 67%, then residual and stacked hybrids benchmarked on the same split.
Aadhaar Document Classifier · PyTorch Vision Transformer EasyOCR XGBoost scikit-learn
Two-stage pipeline for pre-visa document validation: a pretrained ViT extracts visual embeddings, OCR text becomes handcrafted features, and XGBoost classifies on the concatenation.
I write on Medium about things I had to figure out the hard way.
- The day RSA stops feeling unbreakable, and why we should act before it happens · quantum computing, Shor's algorithm, and the case for migrating to post-quantum cryptography now
- Wiring a filter into a living pipeline: my GSoC 2026 final chapter with OWASP OpenCRE · shipping Module B into a production ingestion pipeline
Languages ·
Backend & Frontend ·
ML & Data ·
Cloud & Ops ·
Got a backend that needs to scale, or an inbox that needs to stay private? · Let's talk
In a race to try everything and explore. Star anything you find useful.