SHARVIN NEVE

彷徨の道 // Systems Software & AI Infrastructure Engineer

RTEMS Space Kernel Contributor · Distributed LLM Inference · Delta-RoPE / AgentKV · Edge Fleet Telemetry

GitHub  •  LinkedIn  •  Email  •  Download CV (PDF)

"Preoccupied with a single leaf, you will not see the tree. Preoccupied with a single tree, you will miss the entire forest."
"Do not seek to follow in the footsteps of the wise; seek what they sought."
— Miyamoto Musashi, The Book of Five Rings / Takehiko Inoue's Vagabond

Rather than remaining at the superficial layer of high-level API abstractions and cookie-cutter frameworks, I engineer downward into the bedrock: operating system kernel timers, GPU memory hierarchies, KV cache geometry, and low-latency edge daemons under strict physical constraints.

Mathematical elimination of redundant prefill compute in collaborative multi-agent swarms.

• The Multi-Agent Flaw: State-of-the-art inference engines (vLLM, SGLang) rely on Radix Tree Automatic Prefix Caching (APC), which requires exact prompt alignment from token position 0. In multi-agent swarms sharing large tool specifications, differing agent prefixes trigger an immediate 0% cache hit, forcing full multi-second prefills.
• Delta-RoPE Invariant: Exploits the group-additive properties of Rotary Positional Embeddings: Key vectors obey 2D orthogonal rotation, while Value vectors are position-invariant. Shifting a cached KV block across sequence positions requires only 4 FLOPs per coordinate pair via a single 2D Givens rotation.
• RTX 3060 12GB Benchmark Proof (Qwen 2.5 7B, 1,500 shared tokens):
  • Compute Cost: Slashed from 22.118 TFLOPs to 147.456 MFLOPs (149,995x fewer FLOPs).
  • Rotation Kernel Execution: 0.28 ms vs ~1,105.9 ms (3,949x faster).
  • Analytical Precision: 7.257e-13 max Float64 error (bit-for-bit exact at machine precision, 0.999931 attention cosine similarity).
  • Multi-Agent Pipeline: 3.23x overall pipeline speedup (7.08x TTFT speedup) across a 5-agent collaborative workflow.

🔗 Explore Repository & Math Proof · python verify_rope.py

Empirical throughput scaling and memory-bounded serving harness for open-weights LLMs.

• 10.35x Throughput Scaling: Scaled Qwen 2.5 7B from 57.2 tok/s to 592.0 tok/s across 20 concurrent request streams on an Ampere 12GB GPU.
• PagedAttention & Marlin GEMM: Deployed AutoAWQ Marlin W4A16 GEMM kernels to reserve 3.6 GiB VRAM for a 67,440-token PagedAttention KV-cache pool.
• Latency Bounding: Sustained an 87.1% Radix-Tree Prefix Cache hit rate, reducing P50 Time-To-First-Token (TTFT) to 38.7 ms.

Industrial edge streaming, sensor acquisition, and AI safety platform on ARM/Raspberry Pi.

• Sub-300ms Glass-to-Glass Latency: Hardware-accelerated H.264 video ingestion feeding an embedded Simple Realtime Server (SRS) Docker container, delivered via HTTP-FLV (mpegts.js) with two-way WebRTC audio talkback to a Next.js 16 / React 19 control plane.
• Offline-First & Auto-Chunking: GPIO button triggers 5-minute segmented recording (ffmpeg) with automated sequential cloud synchronization and headless camera-based QR Wi-Fi provisioning.
• Zero-Trust Network Perimeter: Outbound Cloudflare Tunnels (cloudflared) bypassing carrier-grade NAT across 4 industrial field sites with zero open inbound firewall ports.

🔗 Explore Streaming-Rpi Repo · v1.0.0 Release

• End-to-end predictive pipeline with 10 cross-validated models and a Voting Classifier ensemble achieving 84.93% ROC-AUC.
• High-concurrency FastAPI backend with interactive React & Streamlit feature importance visualizers.

Patches to space-grade operating systems, foundation inference runtimes, and Linux primitives:

(May 2026 – July 2026)

• Distributed Edge Cluster: Architected a headless telemetry and video ingestion pipeline across a 6-node ARM edge cluster at 4 industrial sites, processing 18GB of multimodal footage and 14,750+ telemetry events.
• Low-Footprint C++ Daemon: Engineered an edge supervisor daemon with a <45MB RAM footprint, circular memory ring buffers, and systemd watchdog supervision, eliminating crash-loops and recovering cleanly from network dropouts.
• Low-Latency Streaming: Built a sub-300ms WebRTC pipeline (LiveKit / SRS) with hardware H.264 encoding for automated remote aerial inspection, dual-track recording, and real-time visual telemetry capture.

(Aug 2025 – Dec 2025)

• Sub-5ms Sensor Acquisition: Interfaced camera and GPIO/UART/SPI sensors on ARM Cortex boards, hitting sub-5ms acquisition latency.
• Firmware Data Pipelines: Optimized high-throughput circular buffer event serialization in C and Python deployed directly on active field hardware.

• Arbitrary-Position KV Cache Scheduling: Generalizing Delta-RoPE Givens transformations across multi-head and grouped-query attention in distributed inference clusters.
• Space-Grade Kernel Verification: Formal timing boundaries and deterministic lock-free ring buffers on RTEMS for extreme radiation and deep-space missions.
• Zero-Copy Edge Telemetry: Writing eBPF socket filters for zero-overhead packet telemetry on constrained ARM devices.

• 🥇 Smart India Hackathon (SIH) — National Finalist (Hardware & Autonomous Systems Track)
• ☁️ AWS Academy Graduate — Cloud Architecting (AWS Training Badge, 2026)
• ☁️ AWS Cloud Quest — Cloud Practitioner Verified Badge
• 📜 NPTEL Elite Certificate — Design Practices for Intelligent Product Design (IIT Kanpur)
• 🎓 NMIMS MPSTME, Mumbai — Integrated B.Tech + MBA Tech in Computer Engineering (2023 – 2028, CGPA: 3.4/4.0)