CSE @ IIIT Delhi · New Delhi, India

Anushk Kumar

Building practical AI and systems software.

Software Developer (He/Him)

LOCATIONNew Delhi, India

LOCAL TIMELoading...

DIRECT CONTACTanushk3984@gmail.com

QUICK LINKS

Anushk Kumar — Computer Science undergraduate at IIIT Delhi

03 // ABOUT

Building software from
the hardware boundary up.

Computer Science undergraduate at Indraprastha Institute of Information Technology, Delhi (IIIT Delhi). I build practical AI and systems software where performance, privacy, and clear interfaces matter.

04 // GITHUB ACTIVITY

The work leaves a
verifiable trace.

Follow on GitHub

Public Repositories24

Merged External PRs12

Contributions in 2026134

05 // CORE DISCIPLINES

High-signal engineering
domains of practice.

01Close to the silicon

Systems & Compute

High-performance parallel kernels, memory models, hardware instruction simulation, and systems-level memory management.

CUDACC++RustRISC-V

02Models & inference

Applied AI

Deep learning pipelines, multimodal on-device intelligence, quantizing LLMs for CPU HPC, and multi-view cross attention.

PythonComputer VisionNLPGemma

03End-to-end architectures

Product Engineering

Modern full-stack web platforms, native Android & desktop client apps, high-concurrency microservices, and reactive UX.

TypeScriptNext.jsPostgreSQLKotlin

04Provable boundaries

Security

Zero-knowledge storage protocols, authenticated AEAD primitives, packet inspection, and fine-grained role-based policies.

EncryptionNetwork Analysislibpcap

06 // TECH STACK

Technologies I’ve
worked with.

Verified tools, runtimes, and frameworks derived from production codebases, research repos, and course implementations.

Hover or focus any technology to inspect where and how it was used in research & systems.

Python

C++

Rust

CUDA

TypeScript

JavaScript

Kotlin

Java

Next.js

React

React Native

Tailwind CSS

Tauri

Jetpack Compose

Monaco / AceEditor

FastAPI

Spring Boot

Hono

PyTorch

ONNX Runtime

Gemma

Computer Vision

NLP

Vosk

ROS1 / ROS2

ROSBag

OpenMP

RISC-V

PostgreSQL

Drizzle ORM

Better Auth

Supabase

SQLCipher

S3 Storage

Docker

GitHub Actions

Google Cloud / GCP

Vertex AI

LangSmith

Cloudflare CDN

Vercel

libsodium

XChaCha20-Poly1305

libpcap

Figma

Framer Motion

Adobe Illustrator

LaTeX / TeX Live

Three.js

Motion

07 // EXPERIENCE

Experience
that compounds.

Product Intern (Summer)

Adobe Systems · Adobe Illustrator Team

May 2026 – July 2026Noida, India

Engineered an autonomous agent system for direct native desktop application UI control and QE automation.

~3× Workflow Speedup93% Validated TasksPatent in Review Status Method in review for patent

Developed a novel method for controlling core native desktop application UI (Adobe Illustrator) through coordinated agents.
Engineered an autonomous QE testing harness driven by subagents with multi-client socket execution.
Achieved ~3× acceleration in testing workflow speed with 93% validated target task completion.

Undergraduate Researcher (Short Term)

IRAS-Hub · Robotics Research Lab

Jan 2026 – Apr 2026New Delhi, India

Built natural-language retrieval systems for ROSBag robotics telemetry and autonomous vehicle vision pipelines.

ROSBag NLI FocusIIITD-ALIVE ProjectROS1 / ROS2 Platform

Developed a Natural Language Interface (NLI) to query and extract structured telemetry from ROSBag files, streamlining robotics log analysis.
Executed complex simulations and computer vision pipelines to validate autonomous vehicle algorithms for the IIITD-ALIVE project.
Deployed containerized research environments for reproducible execution across heterogeneous robotics hardware.

Undergraduate Researcher (Short Term)

HiPeC — IIIT Delhi · High Performance Computing Lab

July 2025 – Dec 2025New Delhi, India

Quantized and optimized large language models for resource-constrained high-performance CPU architectures.

87.25% HPCorpus AccuracyGGUF + LoRA CompressionONNX + OpenMP Runtime

Engineered a quantized LLM optimized for HPC CPU architectures, drastically reducing inference cost while preserving reasoning accuracy.
Achieved 87.25% accuracy on the HPCorpus dataset by integrating ONNX Runtime, Grid-Ada LoRA fine-tuning, and GGUF compression.
Authored technical documentation detailing the integration of OpenMP-based parallel problem solving with low-resource inference strategies.

Founding Engineer

Resxiv · Integrated Research Support Platform

Jan 2025 – July 2025New Delhi, India

Architected high-availability microservices and a browser-native research platform supporting 100+ concurrent researchers.

100+ Users ConcurrencyMicroservices ArchitectureDocker + CDN Delivery

Architected a scalable backend supporting 100+ concurrent users, utilizing FastAPI and Spring Boot for high-availability microservices.
Developed a reactive frontend with isolated runtime development environments integrating Monaco/AceEditor for code-native capabilities.
Optimized platform SEO via semantic context mapping and automated sitemap generation for low-latency indexing.
Orchestrated automated CI/CD pipelines and cloud infrastructure using Docker and Cloudflare CDN.

08 // PUBLICATIONS & RESEARCH

Peer-reviewed research
& preprints.

2 Research Artifacts

Rigorous empirical benchmarks and multistage attention models tested on thousands of medical studies and forensic video benchmarks.

MAD'25 @ ACM ICMR 2025

SocialDF: Benchmark Dataset and Detection Model for Mitigating Harmful Deepfake Content on Social Media Platforms

Anushk Kumar et al.

A real-world social-media deepfake benchmark combining multimodal face, speech, and contextual verification for forensic authenticity detection.

Total Videos2,126

Benchmark Split1,071 Genuine · 1,055 Manipulated

VerificationMultimodal (Face, Audio, Context)

Read paper

arXiv Preprint (arXiv:2503.01634)Demonstrated at CVPR Demo 2025 (Nashville, USA)

M-SCAN: A Multistage Framework for Lumbar Spinal Canal Stenosis Grading Using Multi-View Cross Attention

Arnesh Batra, Arush Gumber, Anushk Kumar

Automated lumbar spinal canal stenosis grading from cross-sectional 3D MRI series via sequence-based multi-view cross-attention.

Dataset Cohort1,975 Studies

MRI ModalitiesAxial T2 · Sagittal T1 · Sagittal T2/STIR

AUROC Metric0.971

Read preprint

09 // SELECTED PROJECTS

Featured engineering
case studies.

Curated studies in offline multimodal intelligence, zero-knowledge storage, and parallel video processors.

Private Multimodal AI100% On-Device

Amber

Offline-first Gemma-powered Android companion assisting people with dementia through on-device multimodal intelligence.

KotlinGemmaAndroidJetpack Compose

Case study ↗Source ↗

Usable Security & StorageXChaCha20-Poly1305

KryptVault

Zero-knowledge encrypted storage platform utilizing XChaCha20-Poly1305 AEAD and fine-grained access policies.

RustTauriTypeScriptHono

Case study ↗Source ↗

GPU Compute & Kernel EngineeringNVIDIA CUDA

CUDA Parallel Video Processor

High-throughput CUDA-accelerated video filtering pipeline executing custom 2D convolution and spatial kernels.

CUDAC++CMakeDocker

Case study ↗Source ↗

Systems & Computer ArchitectureRISC-V 32-bit

RISC-V 32-bit Assembler & Simulator

High-efficiency 32-bit RISC-V instruction parser and assembler in Rust with a cycle-accurate Python execution simulator.

RustPythonRISC-VComputer Architecture

Case study ↗Source ↗

NLP & Vision RedactionText · Image · PDF · Video

TrueSight

NLP and computer vision tool allowing user-defined gradational masking across text, images, PDFs, and video formats.

PythonComputer VisionNLPOCR

Case study ↗Source ↗

Healthcare EngineeringPostgreSQL + Drizzle

MedicaMS

Clinic management platform with multi-tenant role-based access and an embedded custom database query editor.

Next.jsTypeScriptPostgreSQLDrizzle ORM

Case study ↗Source ↗

10 // HONORS & CERTIFICATIONS

Validated credentials
& national awards.

Specializations

The Johns Hopkins UniversityVerify ↗

GPU Programming Specialization

Advanced multi-course sequence covering parallel hardware architectures, thread scheduling, memory coalescing, CUDA kernel development, and high-performance numerical algorithms.

ID: GP1Z7IRWMP0Y

GoogleVerify ↗

Google Cybersecurity Specialization

Comprehensive professional curriculum spanning network defense, threat intelligence, packet analysis with libpcap/Wireshark, Python security automation, and defensive system architectures.

ID: CKYBKZWMKLPB

Honors & Recognitions

Winner2025

Stage 1 Winner — Grand AI Challenge 2025

NCIIPC / NTRO, Government of India

Awarded 1st Stage Winner in the prestigious national Grand AI Challenge organized by the National Critical Information Infrastructure Protection Centre (NCIIPC / NTRO).

Runner UpDec 2024

Runner Up — Smart India Hackathon (SIH) 2024

Ministry of Education & AICTE, Government of India

Team-level Runner Up in the nationwide Grand Finale of Smart India Hackathon 2024, developing solutions for complex governmental problems.

Premium+2024

Google Arcade Facilitator — Premium+ Milestone

Google Cloud

Achieved Premium+ top milestone facilitating and architecting hands-on labs across Google Cloud Platform, computing infrastructure, and generative AI services.

Demonstration2025

Research Demonstration at CVPR 2025

CVPR Demo Session (Nashville, USA)

Demonstrated M-SCAN multistage spinal stenosis grading framework live at CVPR 2025 demo session.

11 // ACADEMIC FOUNDATION

Indraprastha Institute of Information Technology, Delhi

B.Tech in Computer Science and Engineering

New Delhi, India · 2023 – Present

Computer VisionSystems ProgrammingAlgorithms & ComplexityOperating Systems

12 // RECENT WRITING

Notes on systems
and applied AI.

View all essays

Systems & AI6 min read

Quantizing LLMs for Low-Resource HPC CPUs: Lessons from HiPeC

How we achieved 87.25% accuracy on HPCorpus using ONNX Runtime, Grid-Ada LoRA fine-tuning, and GGUF compression on CPU-only cluster nodes.

Read Essay

GPU Compute5 min read

Writing Custom CUDA Kernels for Real-Time Parallel Video Filters

Deep dive into 2D shared memory tiling, thread divergence mitigation, and CUDA stream pipelining for high-throughput video processing.

Read Essay

Anushk Kumar

Building software from the hardware boundary up.

The work leaves a verifiable trace.

High-signal engineering domains of practice.

Systems & Compute

Applied AI

Product Engineering

Security

Technologies I’ve worked with.

Experience that compounds.

Peer-reviewed research & preprints.

SocialDF: Benchmark Dataset and Detection Model for Mitigating Harmful Deepfake Content on Social Media Platforms

M-SCAN: A Multistage Framework for Lumbar Spinal Canal Stenosis Grading Using Multi-View Cross Attention

Featured engineering case studies.

Validated credentials & national awards.

Specializations

GPU Programming Specialization

Google Cybersecurity Specialization

Honors & Recognitions

Stage 1 Winner — Grand AI Challenge 2025

Runner Up — Smart India Hackathon (SIH) 2024

Google Arcade Facilitator — Premium+ Milestone

Research Demonstration at CVPR 2025

Indraprastha Institute of Information Technology, Delhi

Notes on systems and applied AI.

Building software from
the hardware boundary up.

The work leaves a
verifiable trace.

High-signal engineering
domains of practice.

Technologies I’ve
worked with.

Experience
that compounds.

Peer-reviewed research
& preprints.

Featured engineering
case studies.

Validated credentials
& national awards.

Notes on systems
and applied AI.