Performance-focused Computer Science undergraduate (B.Tech, VIT) specializing in High-Performance GPU Inference, Distributed Systems, Edge AI, and Kernel-Level Sandboxing. Experienced in engineering production systems from first principles—including speculative LLM inference servers, distributed Raft consensus engines, zero-copy video analytics pipelines, and OS-isolated code execution runtimes.
Proven track record of improving data pipeline throughput by 35% at SAIL (Steel Authority of India Ltd.) with a strong command of low-level memory, systems concurrency, and strict hardware budgets.
- Systems & OS Internals: Engineering low-latency runtime kernels, lock-free concurrency pipelines, Windows Kernel Job Object sandboxing, zero-copy memory-mapped I/O (
mmap), and high-performance Raft commit logs. - AI Systems & GPU Acceleration: Specializing in memory-constrained LLM acceleration (KV-cache management, dynamic entropy drafting, speculative decoding), TensorRT dynamic batching, and edge computer vision.
- Distributed Backends: Architecting binary framing protocols over raw TCP, Raft election failovers (<150ms), resilient event-driven microservices, and high-throughput telemetry pipelines.
- Systems & Software Engineering Roles (Distributed Systems, Low-Latency Backends, Core OS)
- Machine Learning Systems & GPU AI Infrastructure Engineering
- High-Impact Open-Source Systems & Research Collaboration
| Domain | Proficiency | Technical Focus |
|---|---|---|
| Speculative LLM Serving | Expert |
Dynamic draft-length adaptation (SVIP), |
| Distributed Systems & Consensus | Advanced |
Raft consensus algorithm, Write-Ahead Logs (WAL), zero-copy mmap binary indexing, raw TCP framing protocols |
| Kernel & Sandboxing | Advanced |
Win32 Kernel Job Objects, restricted token sandboxing, memory quotas, process isolation, sub-25ms cold starts |
| Edge Vision & Acceleration | Advanced |
TensorRT FP16 dynamic batching, multi-stream RTSP decoding (60+ FPS), zero-copy ring buffers, WebRTC sub-80ms streaming |
| Data Pipelines & Telemetry | Advanced |
High-frequency telemetry extraction, ETL pipeline benchmarking, legacy SQL optimization (35% throughput gain at SAIL) |
01. Adaptive Speculative LLM Serving Engine — High-Throughput Local LLM Inference Engine
Architected a high-throughput local LLM inference server implementing entropy-guided dynamic draft length (SVIP) to halt speculative drafting early on uncertain tokens, cutting wasted compute by 45%.
| Dimension | Specification |
|---|---|
| Stack | Python, PyTorch, CUDA, Llama-3.1-8B, Qwen2.5-0.5B, FastAPI |
| Architecture | Dynamic entropy-guided draft gating + |
| Performance | 2.5x token generation speedup over standard autoregressive decoding baselines |
| Memory Bound | Fully contained within an 8GB VRAM envelope without tensor reallocations |
| Impact | Mathematically lossless verification ($\alpha = \min(1, P_{\text{target}}/P_{\text{draft}})$) on consumer hardware |
| Repository | G1Z2P8I7/Speculative-LLM-serving-engine |
Retains KV state directly on device memory, eliminating CPU-GPU serialization bottlenecks and pruning discarded candidate branches in constant time.
02. KhorosLog: Distributed Fault-Tolerant Commit Log Engine — High-Throughput Raft Event Streaming
Engineered a distributed write-ahead log (WAL) and event streaming engine from scratch in Go, implementing the Raft consensus algorithm with randomized election timers and automated leader failover in <150ms.
| Dimension | Specification |
|---|---|
| Stack | Go (Golang), Raft Protocol, Win32 mmap, Raw TCP, CRC32 Checksums |
| Throughput | 80,000+ durable writes/sec bypassing user-space buffer copies |
| Storage Engine | Zero-copy memory-mapped file I/O (mmap) with binary sparse indexing |
| Wire Protocol | Custom framing binary protocol over raw TCP (10x faster serialization than JSON/HTTP) |
| Fault Tolerance | Strict quorum enforcement, persistent log compaction, snapshot replication |
| Repository | G1Z2P8I7/khoros-log |
Designed to eliminate runtime garbage collection overhead on hot commit paths with pooled byte slices and lock-free ring buffer queues.
03. AxiomBox: Kernel-Isolated Remote Code Execution Sandbox — Sub-25ms Secure Runtime Isolation
Engineered a multi-tenant untrusted code execution sandbox achieving sub-25ms cold-start instantiation, outperforming traditional Docker container spin-up latencies by 40x.
| Dimension | Specification |
|---|---|
| Stack | Go, C++, Win32 Kernel APIs, Job Objects, Security Tokens, WebSockets |
| Cold Start | <25ms container-equivalent instantiation latency |
| Isolation | Windows Kernel Job Objects enforcing strict 64MB RAM and 1 CPU core hardware limits |
| Defense | Neutralizes fork-bombs, thread bombs, and memory-exhaustion exploits |
| Telemetry | Real-time kernel accounting streaming stdout, stderr, and memory usage via WebSockets |
| Repository | G1Z2P8I7/Axiombox-Engine |
Provides deterministic sandboxing for online judges, untrusted script evaluation, and interactive code playgrounds with instant teardown.
04. AegisVision: Real-Time Multi-Stream Edge Vision Pipeline — Zero-Copy Video Analytics
High-throughput computer vision engine processing 6 concurrent 1080p RTSP video feeds at 60+ FPS without frame drops by eliminating Python GIL bottlenecks via Win32 Shared Memory (mmap) zero-copy ring buffers.
| Dimension | Specification |
|---|---|
| Stack | C++, Python, TensorRT, OpenCV, WebRTC (aiortc), Win32 IPC |
| Inference | NVIDIA TensorRT FP16 dynamic batching on an RTX 4060 GPU |
| Tracking | ByteTrack spatial geofencing for real-time industrial safety zone intrusion detection |
| Latency | Hardware-accelerated WebRTC streaming delivering sub-80ms glass-to-glass latency |
| Architecture | Decoupled C++ frame grabber + asynchronous Python inference workers over shared memory |
| Repository | G1Z2P8I7/AegisVision |
Eliminates video decoding bottlenecks and provides ultra-low latency spatial telemetry for critical physical security perimeters.
05. Society-Maintenance — Enterprise Apartment Management Platform
Full-stack distributed apartment community management platform orchestrating complaint tracking, financial billing cycles, audit logging, role-based access management, and automated notifications.
| Dimension | Specification |
|---|---|
| Stack | TypeScript, Next.js, React, Prisma ORM, PostgreSQL, Tailwind CSS, Nodemailer |
| Features | Multi-tenant unit administration, financial ledger reconciliations, RBAC permissions |
| Performance | Server-side rendered interfaces with sub-100ms API response latency |
| Reliability | Immutable audit trails, automated billing cron pipelines, and grievance escalation |
| Repository | G1Z2P8I7/Society-Maintenance |
Bokaro Steel City, Jharkhand, India | June 2024 – August 2024
- Architected and benchmarked optimized Python data extraction pipelines to process high-frequency industrial operational telemetry, boosting end-to-end data processing throughput by 35%.
- Automated data integrity validation workflows across cross-departmental internal datasets, cutting manual engineering auditing overhead by 40%.
- Collaborated with senior engineering teams to refactor legacy SQL queries and data models, eliminating database ingestion bottlenecks across heavy operational logs.
#Python #SQL #DataPipelines #Telemetry #DatabaseOptimization #ETL #Automation
2023 – Present
- Architected Adaptive Speculative LLM Serving Engine, achieving a 2.5x speedup on local consumer GPUs via entropy-guided drafting and
$O(1)$ flash KV-cache rollback. - Designed and built KhorosLog, a zero-copy distributed Raft commit log engine in Go processing 80,000+ durable writes/sec.
- Engineered AxiomBox, a lightweight code execution sandbox achieving sub-25ms instantiation via Windows Kernel Job Objects.
#Go #Python #C++ #CUDA #Raft #Win32 #SystemsProgramming
Vellore, Tamil Nadu, India | 2023 – Expected 2027
Bachelor of Technology (B.Tech) in Computer Science & Engineering | CGPA: 7.81 / 10
- Core Coursework: Data Structures & Algorithms, Operating Systems, Database Management Systems, Computer Networks, Object-Oriented Programming, Theory of Computation, Computer Architecture.
- Focus: High-Performance Systems, Low-Latency Networking, GPU Acceleration & LLM Inference, Distributed Consensus.
| Area | Achievement |
|---|---|
| Algorithmic Problem Solving | Solved 250+ data structures and algorithm challenges across LeetCode and CodeChef, focusing on graphs, dynamic programming, and systems concurrency. |
| Industrial Impact (SAIL) | Architected telemetry extraction pipelines yielding a 35% increase in throughput and cutting manual auditing by 40%. |
| Open-Source Systems | Built and published high-performance implementations of speculative decoding runtimes and distributed Raft consensus engines. |
| Campus Leadership (VIT) | Active Member of the Event Management Club (VIT); coordinated logistics and technical staging for major campus events with 500+ attendees. |
engineer:
name: "Sumit Gupta"
handle: "G1Z2P8I7"
role: "Systems & AI Infrastructure Engineer"
education: "B.Tech in Computer Science & Engineering, VIT (2023 - 2027)"
location: "Bokaro Steel City / Vellore, India"
learning:
- "Custom CUDA Kernels for Attention Verification & KV Cache Paging"
- "eBPF Tracing for Low-Latency Linux Socket Monitoring"
- "Distributed Consensus Protocols in Go (Raft, Multi-Paxos Variants)"
building:
- "Adaptive Speculative LLM Serving Engine with Dynamic SVIP Drafting"
- "KhorosLog: Zero-Copy Distributed Commit Log Engine in Go"
- "AxiomBox: Kernel-Isolated Remote Code Execution Sandbox"
exploring:
- "Hardware-Accelerated KV-Cache Compaction Strategies"
- "Zero-Copy Memory-Mapped IPC for Multi-Stream Edge Vision"
- "Kernel Resource Quotas and Containerization Internals"
open_to:
- "Systems & Software Engineering Roles / Internships"
- "AI Infrastructure & GPU Systems Engineering"
- "Core Open Source Systems Collaborations"