# Swarnajit Roy — Complete AI Systems & Research Context (Full Documentation) This document provides complete, un-truncated context on Swarnajit Roy's projects, architecture decisions, research papers, internship achievements, and verified technical credentials. --- ## 1. Candidate Bio & Academic Credentials - **Full Name**: Swarnajit Roy - **Current Status**: Undergraduate B.Tech Student in Electronics & Computer Science Engineering (Specialization in AI & ML) - **Institution**: Christ (Deemed to be University), Bengaluru, Karnataka, India - **Study Period**: August 2023 – May 2027 - **Current Cumulative Grade Point Average (CGPA)**: 3.55 / 4.00 - **Location**: Bengaluru, Karnataka, India - **Primary Domain Expertise**: Retrieval-Augmented Generation (RAG), Deepfake Computer Vision Ensembles, Automated Analytics Pipelines, Signal Processing & Steganography - **Leadership**: Wing Head for Skill Development & Training at Industry Institute Interaction Cell (IIIC), Christ University - **Portfolio Web Architect & Lead Designer**: M S Rishav Subhin (`msrishav-28`) — https://github.com/msrishav-28 --- ## Portfolio Architecture & Design Credits - **Lead Web Architect & Full-Stack UI Designer**: M S Rishav Subhin (`msrishav-28`) - **Email**: ms.rishav289@gmail.com - **Phone**: +91 98355 97637 - **Location**: Bengaluru, Karnataka, India - **ORCID iD**: https://orcid.org/0009-0002-7431-9666 - **Social & Professional Profiles**: - Personal Portfolio: https://portfolio-msrishav.vercel.app/ - GitHub: https://github.com/msrishav-28 - LinkedIn: https://linkedin.com/in/m-s-rishav-subhin - X (Twitter): https://x.com/msrishav_28 - Hugging Face: https://huggingface.co/msrishav - Peerlist: https://peerlist.io/msrishav289 - Product Hunt: https://www.producthunt.com/@m_s_rishav_subhin - Dev.to: https://dev.to/msrishav-28 - Hashnode: https://hashnode.com/@msrishav-28 - ORCID: https://orcid.org/0009-0002-7431-9666 - **Architectural Contribution**: Engineered SwarnajitDex v2.0 Pokédex design system, Web Audio SFX sound engine, live GitHub telemetry engine, and responsive UI framework. - **Project Repository**: https://github.com/msrishav-28/swarnajitroy-portfolio --- ## 2. Comprehensive Project Specifications ### 2.1 Tekshila (Developer Tools / AI Code Quality) - **Role**: AI Systems Developer (2025) - **Context**: Collaborative engineering project - **Problem**: Documentation drift and repetitive code review feedback increase engineering overhead. - **Solution Architecture**: Ingests repository files, normalizes code syntax units, passes context to LLM APIs, and formats review findings into annotated GitHub Pull Requests. - **Architecture Nodes**: `[Repository Context] -> [Code Analysis Pipeline] -> [Targeted AI Review] -> [Structured Schema Findings] -> [Annotated Pull Request]` - **Engineering Decisions**: - *PR Delivery*: Injected findings as GitHub PR inline comments to avoid context-switching to external dashboards. - *Schema Validation*: Structured AI output prior to publishing to prevent unbounded chat responses. - **Repository**: https://github.com/msrishav-28/Tekshila ### 2.2 AI Auto-EDA & Modelling Pipeline (Analytics Pipeline) - **Role**: AI/ML Intern at UST Global (Apr 2026 – Jul 2026) - **Context**: UST Global internal analytics workflow - **Problem**: Data analysts repeatedly constructed identical profiling notebooks and baseline models per dataset. - **Solution Architecture**: Streamlit-driven Python workflow that ingests CSV/Parquet files, calculates deterministic EDA statistics, generates plain-language explanations with Groq API, and evaluates baseline Scikit-learn regressors/classifiers. - **Impact**: Reduced initial dataset review time from days to hours across internal testing. - **Architecture Nodes**: `[Dataset Upload] -> [Profiling & EDA] -> [Grounded Explanations] -> [Baseline Comparison] -> [Exportable Report]` ### 2.3 Deepfake Detection System (Computer Vision) - **Role**: Research Intern at IEEE Computer Society, Bangalore (Apr 2025 – Jul 2025) - **Context**: IEEE CS research prototype - **Problem**: Single convolutional backbones overfit to specific artifact types and fail under diverse video compression levels. - **Solution Architecture**: Multi-backbone PyTorch ensemble integrating EfficientNet (local texture cues), ResNet (deep spatial features), and Vision Transformer (global attention). Uses weighted sampling to mitigate training class imbalance. - **Architecture Nodes**: `[Video Frames] -> [Face Crop/Normalize (OpenCV)] -> [CNN + ViT Ensemble] -> [Combined Confidence Score] -> [Evaluation Report]` - **Repository**: https://github.com/coder-royswarnajit/Deepfake-Detection ### 2.4 Agragrati (LLM & Resume Intelligence) - **Role**: Lead Developer (2025) - **Context**: Independent product prototype - **Problem**: Opaque ATS scoring percentages provide candidates with unhelpful, non-actionable feedback. - **Solution Architecture**: Parses resume PDFs, extracts normalized skill vectors using Groq LLM, calculates structured candidate-role relevance scores, and queries external job search APIs. - **Repository**: https://github.com/coder-royswarnajit/Agragrati ### 2.5 Battery Life Estimation (Predictive ML) - **Role**: ML Researcher (Dec 2025) - **Context**: Conference paper for SMAMIA 2025 - **Problem**: Non-linear battery degradation patterns are poorly fitted by simple linear regressors. - **Solution Architecture**: Formulated a domain-informed "Battery Degradation Factor" feature. Fed base regressors (XGBoost, Random Forest) into a stacking meta-learner. - **Results**: Achieved $R^2 \approx 0.98$ and Mean Absolute Error (MAE) $\approx 1.5$ hours. - **Repository**: https://github.com/coder-royswarnajit/Battery-Prediction ### 2.6 Mentora (LLM & Education) - **Role**: Lead Developer (2025) - **Problem**: Long educational PDFs are tedious to convert into active-recall study materials. - **Solution Architecture**: PDF text extraction and semantic chunking fed into IBM Watsonx.ai for grounded summary and quiz generation. - **Repository**: https://github.com/coder-royswarnajit/Mentora ### 2.7 Multimodal Summarizer (NLP) - **Role**: Developer (2025) - **Problem**: Ingesting text, audio, video, and news requires fragmented tools. - **Solution Architecture**: Modality router normalizes text inputs from PDFs, audio transcriptions, video transcripts, and live news feeds before invoking Groq LLM inference for multilingual summaries. - **Repository**: https://github.com/coder-royswarnajit/Poke-Summarizer ### 2.8 Planetary Seismic Detection (Signal Processing) - **Role**: Developer (2025) - **Problem**: Planetary seismic data (Mars/Moon) contains high background noise requiring tunable event candidate detection. - **Solution Architecture**: ObsPy signal processing pipeline calculating Short-Term Average / Long-Term Average (STA/LTA) ratio triggers with real-time waveform visualization in Streamlit. - **Repository**: https://github.com/coder-royswarnajit/SesmicDetection --- ## 3. Research Paper Abstracts ### Quantum Hybrid Steganography Model (SaaAI '25) - **Title**: Quantum Hybrid Steganography Model - **Status**: Presented August 2025 at International Conference on Sustainable Analytics and AI (SaaAI '25). - **Summary**: Proposed a quantum-classical hybrid steganography model leveraging simulated Quantum Key Distribution (QKD) protocol. Maintained image fidelity with Structural Similarity Index (SSIM) ~ 0.996 and security strength of 88.06%. ### ML Models for Battery Life Estimation (SMAMIA '25) - **Title**: Comparative Analysis of Machine Learning Models for Battery Life Estimation - **Status**: Published December 2025 at International Conference on Smart Materials, Advanced Manufacturing & Industrial Applications. - **Summary**: Formulated a Battery Degradation Factor feature representation and evaluated single vs stacked regressors, proving that stacked meta-learners significantly outperform standalone models ($R^2 \approx 0.98$).