Rahul Paul — Portfolio

9.52/10CGPA · CSE (Cyber Security)

Top 7.3%LeetCode · rating 1821

6.7×on-device model compression

100%precision & recall · Sentinel benchmark

1,500+cubes solved · 100% correct

About

~/about

Most of what I know, I've learned by building. I like taking a hard problem end to end and staying with it through the unglamorous parts — the tests, the CI, and the bug that quietly halves your scores. My projects tend to sit where machine learning, web engineering, and security meet: a Rubik's-cube solver held to a near-optimal move count, a DNS threat detector built to read real attack traffic, and an on-device captioning model squeezed into 153 MB. I care about shipping things that actually run, not just demos. Right now I'm a project intern at Samsung R&D, and I'm looking for internships where I can keep learning from people who build at scale.

// what I'm looking for

Open to software engineering, machine learning, quantitative, and security internships in technology and finance. Graduating 2027 · based in Bengaluru, India. Fastest way to reach me: email.

MS Ramaiah Institute of Technology

B.E. in Computer Science and Engineering (Cyber Security)

Bengaluru, Karnataka
2023 – 2027 · CGPA 9.52 / 10

Bhavan's Gangabux Kanoria Vidyamandir

Class XII (AISSCE)

Kolkata, West Bengal
2021 – 2022 · 96.33%

Experience

~/experience

Project Intern · Samsung PRISM

Dec 2025 – Present

Samsung R&D Institute India · Bengaluru

Developing EmoCapNet, a 254.8M-parameter multimodal vision-language model for 7-class emotion-aware image captioning — fusing ViT-Base + GPT-2 via cross-attention and VAD-FiLM conditioning, trained with multi-task learning (captioning, VAD regression, emotion classification, contrastive alignment) on a 67K-caption emotion-augmented Flickr30k corpus.
Distilled the teacher into a TinyCLIP-ViT + DistilGPT-2 student (112.5M params, 2.3× smaller) via soft-logit, VAD/emotion-head, and feature-level embedding distillation; measured METEOR 22.1, CIDEr 0.363, and an emotion-controllability correlation of r = 0.81 on a 6.7K-caption held-out set.
Engineered on-device deployment via INT8 dynamic quantization (custom Conv1D→Linear rewrite to make GPT-2 quantizable), compressing ~1GB → 153MB (6.7×); built greedy KV-cache decoding from scratch (524 ms/caption on CPU) with an ONNX/TorchScript export path.

Projects

~/projects · ls

github.com/rahulpaul-07/sentinelcode ↗

Sentinel Featured

Autonomous AppSec Agent

An AI agent that finds, proves, and fixes vulnerabilities in source code: an LLM hunts candidate bugs, then a validator must prove each one with a proof-of-concept exploit run in a locked-down Docker sandbox — unproven claims are discarded, driving false positives toward zero.
Provider-agnostic model layer (LiteLLM) runs the same pipeline on local Ollama, Claude, or OpenAI; a patcher then proposes secure-fix diffs behind a human-approval gate.
On a small, hand-labeled benchmark across five vulnerability classes (SQL/command/path/deserialization, hardcoded secrets), a self-correction loop raised recall from 60% to 100% with precision holding at 100%, measured by a reproducible eval harness. Inspired by DARPA's AI Cyber Challenge.

PythonLLM AgentsLiteLLMOllama / Claude / OpenAIDocker sandboxASTpytest

View code Write-up

smart-course-generator.vercel.applive ↗

Smart Course Generator

May – Jun 2026

Full-stack AI learning platform that generates structured multi-module courses and streams them lesson-by-lesson over SSE, with adaptive quizzes, flashcards, and an AI-scored mock-interview mode.
Custom multi-provider AI router (Gemini → Groq → OpenRouter) with per-provider circuit breakers, retries, timeouts, and API-key rotation; quality made measurable via an LLM-as-judge eval harness with RAG grounding, wired into CI.
Secured with rotating, reuse-detecting refresh-token auth, Zod validation, Helmet, and rate limiting.

ReactTypeScriptNode.jsExpressMongoDBSSEZodJestPlaywright

Live demo Code

rubiks-cube-studio.vercel.applive ↗

Rubik's Cube Studio

Mar – Apr 2026

Implemented Kociemba's two-phase algorithm from scratch in TypeScript (cubie model, coordinate reduction, precomputed pruning tables, IDA* search) — averaging ~20.6 moves at 100% correctness across 1,500+ random cubes vs. an independent reference engine.
Ran the solver in a Web Worker to drive a stutter-free Three.js 3D animation; added webcam cube scanning via an HSV color pipeline.
Shipped with 66 unit tests, Playwright E2E, strict TypeScript, and CI.

TypeScriptThree.jsWeb WorkersViteVitestPlaywright

Live demo Code

dns-sentinel.vercel.applive ↗

DNSentinel

Apr – Jun 2026

DNS Threat Intelligence Platform

Full-stack DNS threat-detection platform (Manifest V3 Chrome-extension telemetry) classifying DGA, tunneling, exfiltration, and C2 beaconing via a 22-feature Random Forest + Isolation Forest ensemble.
Adaptive risk engine (µ + k·σ thresholding) driving a SOAR auto-block / sinkhole layer mapped to MITRE ATT&CK, streamed to a SOC dashboard over SSE.
Diagnosed a scoring-inversion bug where a disabled model silently halved every ensemble score; validated with a benchmark over 25 real DGA families plus SHAP-explained PDF reports.

PythonFastAPIReactscikit-learnDockerChrome MV3pytest

Live demo Code

More on GitHub view all repos ↗

Rubiks-Cube-Solver

High-performance C++ cube solver using IDA* search, pattern-database heuristics, and a bitboard representation — solves scrambles in under a second (Release build).

C++

multi-source-candidate-transformer

Merges messy, conflicting candidate data from many sources into one canonical, deduplicated profile — provenance and confidence on every field, union-find identity resolution, fully deterministic and tested.

Python

AI-Botnet-Traffic-Detection

Real-time botnet intrusion detector — XGBoost/LightGBM on network flows with live Scapy packet capture, a FastAPI + Streamlit dashboard, SHAP explainability, and Docker.

Python

AI-Help-Extension

Context-aware Chrome extension (Manifest V3) that embeds a Google Gemini AI tutor into the Maang.in CP platform — reads the problem and your code to give specific hints.

JavaScript

Chrome-Extension-Problem-Tracker

Chrome extension (MV3) to bookmark and track competitive-programming problems, backed by the Chrome Storage API.

JavaScript

Writing

~/writing

Applications of Quantum Computing for Internet of Things

Springer Nature · book chapter, pp. 79–121 · 2025

Co-authored a 43-page chapter in Quantum Computing, Sensing and Communications for IoT, surveying 70+ papers on integrating quantum computing with IoT for security, optimization, and intelligent decision-making across Quantum ML, Post-Quantum Cryptography, and Quantum Sensing.

Read on Springer

SOON

Build logs in the works — writing up the DNSentinel scoring-inversion bug and what it took to build a near-optimal Rubik's-cube solver from scratch.

Technical Skills

~/skills · cat

Languages

PythonC / C++TypeScriptJavaScriptSQL

AI / ML & Deep Learning

PyTorchTransformers (ViT, GPT-2)Multimodal VLMsKnowledge DistillationINT8 QuantizationONNXTorchScriptscikit-learn

GenAI & Web

LLM IntegrationRAGPrompt EngineeringMulti-provider AI RoutingReactNode.jsExpressFastAPIREST APIsSSEJWT / OAuth2

Data, DevOps & CS

MongoDBSQLAlchemyDockerGitGitHub Actions (CI/CD)JestVitestPlaywrightpytestDSAOOPDBMSComputer NetworksCyber Security

Beyond the code

~/interests

Competitive programming

I compete on LeetCode (rating 1821, top ~7.3% over 17 rated contests) and like keeping my data-structures and algorithms sharp between builds.

Algorithmic puzzles

My favourite work is often the algorithmic kind — implementing Kociemba's two-phase cube solver from scratch, or reverse-engineering DGA behaviour for a threat detector.

Emerging tech

I enjoy reading and writing where fields collide — I co-authored a Springer survey on quantum computing for IoT, spanning quantum ML and post-quantum cryptography.

Achievements & Certifications

~/credentials

Achievements

1821LeetCode contest rating · top 7.32% globally (17 rated contests)

LeetCode — competitive programming profileprofile ↗
Campus Ambassador — Kshitij, IIT Kharagpurcertificate ↗

Certifications

IBM — Threat Intelligence & Huntingverify ↗
IBM — Security Operations in Practiceverify ↗
Infosys Springboard — Data Structures & Algorithmsverify ↗
MongoDB — Introduction to MongoDBverify ↗
Google Cloud — Career Launchpadverify ↗

Contact

~/contact

Open to software-engineering, machine-learning, quantitative, and security internships in technology and finance (graduating 2027). The fastest way to reach me is email — happy to chat.

rahulpaul.csetech@gmail.com copied ✓ +91 91631 30524 github.com/rahulpaul-07 LinkedIn