sharvin@edge-node:~$ neofetch --profile
OS: Linux / Edge Fleet Architecture (6-node distributed cluster)
Host: NMIMS MPSTME, Mumbai (Integrated B.Tech + MBA Tech '28)
Kernel: Modern C++20, Python, POSIX pthreads, WebRTC
Experience: Software & Systems Intern @ Aspire Consultancy Services | Ex-Intern @ Entice Engineering
Core Focus: Distributed LLM Inference (vLLM, SGLang) & High-Throughput Edge Telemetry
Current State: Architecting offline-first streaming daemons & instrumenting AI infrastructure

I am a computer engineering student and systems software engineer building at the intersection of low-level systems and applied AI infrastructure.

Rather than stopping at high-level API wrappers, I work deep in the stack: contributing latency telemetry upstream to tier-1 LLM inference engines (SGLang, vLLM / FlashAttention), engineering memory-constrained C++ daemons for multi-node edge camera fleets, and shipping full-stack predictive ML platforms. My guiding ethos: if it doesn't survive network partitions and high I/O concurrency, it isn't ready for production.

A selection of upstream pull requests contributed to production AI inference engines and systems infrastructure:

(May 2026 – July 2026)

• Distributed Edge Telemetry: Architected an offline-first distributed telemetry and video data pipeline across a 6-node edge fleet, reliably ingesting 18GB of multimodal footage and 14,750+ telemetry events across intermittent network partitions.
• Low-Footprint C++ Daemon: Engineered an edge supervisor daemon in C++ with a <45MB RAM footprint and systemd watchdog supervision, achieving zero crash-loops and continuous automated cloud re-synchronization.
• Sub-300ms Live Streaming: Integrated WebRTC streaming (LiveKit + FFmpeg + Next.js) and multimodal API ingestion for automated real-time inspection, dual-track recording (local & desktop), and AI snapshot comparison.

(Aug 2025 – Dec 2025)

• High-Throughput Serialization: Built event serialization routines in C and Python, hitting sub-5ms data acquisition latency under sustained high-I/O sensor throughput.
• Hardware Integration: Developed automated hardware integration and telemetry synchronization pipelines deployed across production sensor systems.

Production-grade machine learning platform for customer churn analytics with automated feature transformation and real-time scoring.

• Calibrated Ensemble Model: Engineered an end-to-end predictive pipeline analyzing high-dimensional user interaction data, achieving an 84.93% ROC-AUC with calibrated XGBoost and LightGBM models.
• Full-Stack Architecture: Built a high-concurrency FastAPI backend with interactive React & Streamlit evaluation dashboards for lead triage and feature importance visualizers.

End-to-end edge-to-cloud live video and telemetry pipeline built for mission-critical remote inspection.

• Real-Time WebRTC Streaming: Raspberry Pi camera feed hardware-encoded via FFmpeg and streamed over LiveKit WebRTC with sub-second glass-to-glass latency to a Next.js control dashboard on Vercel.
• Edge Analytics: Integrated local and desktop synchronized recording, two-way audio channels, and edge AI snapshot comparison.

→ Streaming-Rpi Repo · → Centrix-Helmet Repo · → pi-0 Repo

• Distributed KV Cache & Memory Scheduling: Exploring PagedAttention memory layouts and chunked prefill dynamics across distributed inference instances (vLLM / SGLang).
• Agentic Workflows via Model Context Protocol (MCP): Building multi-agent systems with deterministic tool routing, dynamic context pruning, and sandboxed execution boundaries.
• Kernel-Level Observability: Writing eBPF probes for zero-overhead telemetry tracking on edge Linux daemons and low-latency network sockets.

• 🥇 Smart India Hackathon (SIH) — National Finalist (Hardware & Systems Software Track)
• ☁️ AWS Cloud Quest: Cloud Practitioner — Verified Cloud Architecture Credential
• 📜 NPTEL Elite Certificate — Design Practices for Intelligent Product Design (IIT Kanpur)
• 🎸 Musician & Band Member — Practicing creative discipline, live rhythm, and team dynamics