Résumé

Shikhar SaxenaBackend and Forward-Deployed Engineer

Open to work

I build backends and make them fast on real hardware: NodeJs, Go and FastAPI services that keep state consistent under load, LLM inference and vector search tuned on GPUs, and edge systems deployed right next to the cameras they serve. Final-year computer science at Thapar Institute.

Fast is a number, not a feeling.

  • Decode latency

    14–15 ms

    per token, LLM inference under load

  • GPU speedup

    22×

    batch vector search vs CPU

  • Concurrent users

    1000+

    at sub-100 ms on Zenith

  • False alarms cut

    65–90%

    by ZaiGuard's filter layers

  • LeetCode

    417+

    problems solved, 1551 rating

Experience

Measured, then shipped.

Two internships: one tuning LLM inference on GPUs, one building the backend that dispatches a truck fleet.

Jun – Jul 2026

Voxomos Systems

Software Intern

  • Developed and benchmarked CPU, CUDA and HIP implementations of cosine similarity and vector search, profiled with NVIDIA Nsight Systems and Nsight Compute, and tested portability to AMD through ROCm, HIP and HIPIFY.
  • Evaluated inference across vLLM, SGLang and TensorRT-LLM, isolated KV-cache memory limits under load, and cut time-to-first-token spikes by tuning request concurrency.
Decode latency
14–15 ms/token
Batch vector search
22× faster than CPU
Runtimes
vLLM, SGLang, TensorRT-LLM
Platforms
CUDA, ROCm/HIP
Read the internship study
Jan – Apr 2026

TruckHai

Backend Intern

  • Built Go (Gin, GORM) REST microservices for fleet dispatch, modelling the trip lifecycle as a finite-state machine with ACID transactions so invalid transitions are rejected and concurrent updates stay consistent.
  • Removed race conditions and row-lock contention in PostgreSQL with composite indexes and query-plan tuning, and added RBAC middleware with audit-logged admin impersonation.
Stack
Go, Gin, GORM, PostgreSQL
Trip lifecycle
Finite-state machine
User roles
5+ with RBAC

Internship study

Same GPU, three engines.

At Voxomos I served TinyLlama-1.1B on vLLM, SGLang and TensorRT-LLM on one 6 GB RTX 3050, then wrote the GPU kernels behind vector search. These are the numbers that moved.

  • Decode floor

    13.7 ms

    per token at no load; vLLM and SGLang hit 15.1. The GPU sets the floor, not the engine.

  • Admission control

    1.2 s

    to first token with 32 requests in flight, against 25–76 s uncapped.

  • Prefix caching

    2.8×

    more vLLM requests per second when prompts share a 500-token prefix.

  • Energy per token

    44×

    TensorRT-LLM tokens per watt at batch 768 against batch 1.

Every engine and workload, the vector search kernels, and what the numbers add up to.

Read the full study

Projects

Built under constraint.

Three systems, each shaped by one hard problem: noise, trust or traffic.

Surveillance system

65–90%

fewer false alarms, sub-second

ZaiGuard

FastAPI, PostgreSQL, Qdrant, Redis, YOLO, ByteTrack, Flower

  • Edge node reads RTSP, HTTP or USB feeds with motion-gated sampling into YOLO26s/YOLO11s on ONNX (96.5% precision) and ByteTrack.
  • Classifiers flag after-hours access, and dog attacks from lunge acceleration, thrashing and proximity, scored 0–100.
  • Five-layer FastAPI alert engine (below); a transactional outbox keeps PostgreSQL and Qdrant consistent. 101 automated tests across three services.
How it works and the stack
Shikhar1808/Zaiguard-Connected
HackMol 6.0

Voice → FIR

evidence anchored on Ethereum and IPFS

CivicSentinel

MongoDB, Express, React, Node.js, TypeScript, Solidity, LangChain

  • Intake pipeline (Twilio, GPT-4, LangChainJS) that turns unstructured audio and text into validated FIR records with automated severity triage.
  • Evidence hashes anchored on Ethereum smart contracts and IPFS for a tamper-proof audit trail.
  • OTP citizen sign-in, background task queues, and latency and uptime monitoring with Prometheus and Grafana.
How it works and the stack
Shikhar1808/HackMol-6.0Live demo
Competition platform

< 100 ms

responses at 1000+ concurrent users

Zenith

Node.js, Express, MongoDB, JWT, NGINX, AWS

  • Real-time leaderboard calculation and stateless JWT auth to keep the database out of the way during submission spikes.
  • Cloudflare DDoS mitigation plus NGINX reverse-proxy rate limiting and load balancing.
How it works and the stack
Shikhar1808/ZenithCryptic

How ZaiGuard decides on every event

Alert engine, in order

  1. Sensitivity gate

    Layer 1

    Drops detections below a threshold adjusted for the hour and the zone. (Time and zone rules)

  2. Enrichment

    Layer 2

    Attaches camera, zone and a readable description to each event. (Metadata)

  3. Burst de-duplication

    Layer 3

    Merges repeats of an active incident; re-alerts only on a confidence jump. (Redis)

  4. Rule suppression

    Layer 4a

    Silences anything that matches an operator's suppression rule. (PostgreSQL)

  5. Semantic suppression

    Layer 4b

    Silences events too similar to an alert someone already dismissed. (Qdrant)

  6. Priority tiering

    Layer 5

    Grades what survives as critical, high or medium, then publishes it. (Outbox worker)

More on GitHub

All repositories

  • Federated learning hub · Flower, PyTorch

    FedAvg and FedProx on a real SuperLink and SuperNodes across laptops, with non-IID data partitioning.

  • Jetson multi-camera surveillance · Jetson Nano, PyTorch, FFmpeg, Docker

    Four camera feeds with detection, SORT tracking, snapshot alerts with cooldown, and FFmpeg grid recording.

  • PokeCNN · Next.js, FastAPI, PyTorch, ONNX

    A Pokémon-themed playground: snap together conv, activation, batch-norm and dropout blocks, and the backend builds, trains and exports the model.

  • BISWISE · React, Kaboom.js, Tiled

    A retro pixel-art game that teaches BIS quality standards through quests, quizzes and a leaderboard.

  • Token launchpad · Solidity, OpenZeppelin, Hardhat

    Create ERC20 tokens, price them in ETH and trade them, with reentrancy-guarded buy, sell and withdraw, tested in Hardhat.

  • GitHub Discord bot · Discord.js, GitHub API

    Commands that pull a GitHub profile, a repository or a user's full repo list into Discord.

  • BlogWebApp · Express, MongoDB, JWT, Multer

    A blog with cookie-based JWT sign-in, image uploads and server-rendered pages.

  • Contacts manager · Express, MongoDB, JWT, bcrypt

    A REST API where each signed-in user manages only their own contacts.

  • Slack clone · Socket.io, Express

    Real-time chat organised into Socket.io namespaces and rooms.

  • SlowBro · Express 5, MySQL, Oracle

    REST API for a course platform: users, courses, course content and enrolments, built from an ER design.

Stack

The toolbox.

What I've used across internships, projects and hackathons.

  • Languages

    • C++
    • C
    • Go
    • Python
    • JavaScript
    • TypeScript
    • SQL
    • Solidity

  • Backend & DevOps

    • Go
    • Gin
    • GORM
    • FastAPI
    • Microservices
    • REST APIs
    • RBAC
    • PostgreSQL
    • MySQL
    • Redis
    • Docker
    • Kubernetes
    • Helm
    • Linux
    • GitHub Actions
    • NGINX
    • Prometheus
    • Grafana
    • Loki
    • AWS

  • AI & GPU

    • PyTorch
    • CUDA
    • AMD ROCm/HIP
    • Nsight
    • vLLM
    • SGLang
    • TensorRT-LLM
    • ONNX Runtime
    • YOLO
    • ByteTrack
    • OpenCV
    • LangChain
    • LangGraph
    • MCP
    • FAISS
    • Qdrant
    • Flower
    • Jetson

  • Web (MERN)

    • MongoDB
    • Mongoose
    • Express
    • React
    • Node.js
    • Next.js
    • Tailwind CSS
    • TanStack Query
    • React Hook Form
    • Zod
    • Socket.io
    • JWT
    • Discord.js
    • Hardhat
    • OpenZeppelin

Record

On the record.

Education, awards and certifications.

WhatWhereResultWhen
B.E. Computer Science and EngineeringThapar Institute of Engineering and Technology, PatialaCGPA 8.59 / 102023 – 2027
Class 12 and Class 10Army Public School, Bareilly92.4% and 93%2023
SPHNIX'24 HackathonMNIT Jaipur1st runner-upNov 2024
NVIDIA certificationsAccelerated Computing (Modern CUDA C++, CUDA Python), AI on Jetson Nano, Generative AI & Deep LearningCertified2025 – 2026
LeetCodeArrays, graphs, DP, trees, heaps417+ solved, 1551 ratingOngoing

Hiring for backend, forward-deployed or GPU work? Let’s talk.

saxena.shikhar1808@gmail.com

GitHubLinkedInLeetCodeRésumé