Flash Attention implementation, built from scratch in PyTorch, that reduces attention memory complexity from O(N²) to O(N) using tiled computation and online softmax. Includes V1/V2 variants with memory-efficient backward passes, GQA/MQA support for inference optimization, and a custom Triton kernel achieving an 11.8× speedup over naive attention. Validated across 146 tests on A100 and H100 GPUs, demonstrating deep understanding of GPU memory hierarchies and numerical algorithms.

Developed Substrata9, a lightweight systems-diagnostics toolkit built entirely in Bash for deep Linux process introspection. It parses raw /proc filesystem data to reveal memory maps, file descriptors, process hierarchies, and runtime anomalies without requiring compilation or external dependencies. The suite includes modular CLI utilities with JSON output for automation and integrates seamlessly into observability, debugging, and forensics workflows.

Created a Graph Retrieval-Augmented Generation (GraphRAG) web app that answers Mission: Impossible questions with context-aware, generative responses. The system enhances traditional RAG by combining cosine-similarity search on semantic embeddings with a dynamically constructed knowledge graph, enabling deeper contextual understanding. A Flask backend builds and queries the graph using NetworkX, while a language model generates responses based on rich subgraph context. The application runs under Gunicorn (WSGI) and is fronted by NGINX as a reverse proxy, with communication handled via a Unix socket for secure, low-latency performance. Hosted on multi-zone Google Compute Engine, the service leverages GCE’s 99.99% uptime SLA, with tightly scoped ingress rules for high-performance, secure access.

Designed and deployed a secure, real-time log-analytics platform on Google Cloud that ingests, processes, and visualizes network logs end to end. Architected a custom Virtual Private Cloud (VPC) sliced into three /24 subnets—x.y.1.0/24 (web), x.y.2.0/24 (application), and x.y.3.0/24 (processor)—each pinned to dedicated Compute Engine VMs to enforce zero-trust micro-segmentation and east-west isolation. Granular, stateful firewall rules admit traffic only from whitelisted IP prefixes and service accounts. Logs are encrypted in flight over SSH, transformed with Python, staged in Cloud Storage, and streamed through Pub/Sub to invoke Cloud Functions (1st gen.) that load structured data into BigQuery. A hardened Flask API—exposed via HTTPS and IAM-based authentication—delivers controlled, low-latency access to analytics, providing scalable, compliant, and high-performance troubleshooting insights.

Built an interactive visual learning platform (Graphil) using React to simplify complex technical topics—including Linux, GCP, networking, Python, and AI—through modular, pre-rendered visualizations. The intuitive UI/UX design enables self-guided exploration of technical subjects, enhancing comprehension for visual learners. The platform is open-source, encouraging community collaboration and extensibility.