Available for opportunities

I build secure
intelligent systems.

K Harshit — CS undergrad at PES University specializing in LLM security, adversarial ML, and building production-grade AI threat detection pipelines.

Projects

Research-driven builds across LLM security, AI steganography, systems programming, threat intelligence, and production-grade backend systems.

DCASS

AI + Steganography

Dynamic Context-Aware Semantic Steganography — a research system that encodes covert messages by curating semantically aligned media using multi-modal AI embeddings instead of modifying carrier files.

  • Multi-modal embedding pipeline using CLIP, CLAP, and SentenceTransformers for semantic alignment across text, image, and audio.
  • Reinforcement-learning agent (Stable-Baselines3) for optimal carrier selection with FAISS-powered vector search.
PythonPyTorchCLIP/CLAPFAISSRL

XJailGuard

LLM Security

Explainable LLM security framework with modular pipeline for input sanitization, intent classification, and output validation against prompt injection attacks.

  • Modular pipeline with independently configurable stages for input sanitization, intent classification, and output validation.
  • Cross-lingual threat detection supporting 10+ languages by fine-tuning multilingual transformer models on adversarial prompt datasets.
PythonPyTorchNLPSHAP

CTI-RAG

Threat Intel

RAG pipeline over CVE and MITRE ATT&CK datasets with custom chunking strategies and vector indexing for high-precision semantic retrieval of threat intelligence.

  • RAG pipeline over CVE and MITRE ATT&CK datasets with custom document chunking and ChromaDB vector indexing.
  • LangChain-based query engine grounding LLM responses in retrieved threat intelligence, reducing hallucination.
LangChainChromaDBStreamlitPython

SecureLogger

Cyber Deception

GAN-powered adversarial log generation system that synthesizes realistic server access logs to obfuscate attacker behavioral fingerprints in honeypot environments.

  • GAN architecture trained to synthesize realistic server access logs that statistically mirror real traffic distributions.
  • Discriminator trained on real server log datasets to enforce authenticity, iteratively refining Generator outputs.
PyTorchGANFlaskPython

Custom VPN

Systems Programming

Layer 3 VPN tunnel built from scratch using C, raw sockets, and Linux TUN/TAP interfaces with a Docker-based client-server architecture for network isolation.

  • Raw socket programming in C with Linux TUN/TAP interfaces for Layer 3 packet tunneling over UDP.
  • Docker-based client-server architecture with the client isolated in a container for secure testing.
CRaw SocketsLinux/TUN/TAPDocker

Intelligent Research Analysis

NLP + Graphs

Automated ingestion pipeline that parses and semantically embeds academic research papers, then builds a directed knowledge graph for gap analysis.

  • Semantic embedding of research papers using SentenceTransformers for dense vector representations.
  • Directed knowledge graph with NetworkX mapping citation relationships and surfacing contradictions.
PythonFastAPISentenceTransformersNetworkX

Technical Skills

Organized by domain — from core languages to specialized security and ML tooling.

Languages & Core

01

Primary programming languages

PythonJavaCJavaScriptTypeScriptSQL

AI & Machine Learning

02

ML frameworks, NLP, and model training

PyTorchHugging FaceNLPGANsReinforcement LearningSentenceTransformersCLIP / CLAPVector EmbeddingsXAI (SHAP)LangChain

Security & Adversarial

03

Offensive/defensive security and LLM safety

LLM SecurityAdversarial MLThreat ModelingOWASP Top 10PentestingSteganographyJWT Auth

Backend & Data

04

Server frameworks, databases, and APIs

Spring BootFastAPIFlaskNode.jsREST APIsMongoDBChromaDBFAISSNext.js

Systems & Networking

05

Low-level systems and network programming

Socket ProgrammingLinuxBashDockerTUN/TAP InterfacesNmap

Tools & Workflow

06

Development tools and platforms

Git / GitHubStreamlitNetworkX

Experience

Summer Research Intern

CCNCS, PES University

June 2025 - July 2025
  • Designed XJailGuard, a modular LLM security framework to detect multi-turn and cross-lingual prompt injection attacks in transformer-based models.
  • Built multilingual classification pipelines using PyTorch and Hugging Face, incorporating sliding-window contextual memory for adversarial intent detection.
  • Improved jailbreak detection accuracy from 8-12% baseline to 91-93% through layered input/output filtering and zero-trust validation.
  • Integrated SHAP-based Explainable AI and evaluated performance using false positive/negative rates and latency benchmarks; authored for ICICC 2025 publication.
PyTorchHugging FaceSHAPNLP

Education

B.Tech in Computer Science and Engineering

PES University, Karnataka, India

2023 - 2027
8.54CGPA
On track

Achievements

Prof. C N R Rao Merit Scholarship

Awarded for academic excellence at PES University.

Certifications

Cybersecurity - Basics of Red Teaming

Fundamentals of ethical hacking, reconnaissance, exploitation, and post-exploitation techniques.