Girish Kulathumani
Computer Science and Engineering graduate (VIT Chennai, 2025). Research interests: decoding non-invasive biosignals (EEG and wrist surface EMG) so that models trained on some people work on new ones with little or no calibration; personalising shared models without pooling each person's data, using federated learning; and evaluation with subject-wise splits that rules out data leakage.
Work Experience
Application Engineer
Built Salesforce Life Sciences Cloud features (Apex, Lightning Web Components, Flows) for a global pharmaceutical client.
- Evaluated GitHub Copilot on real user stories, reporting over 80% average time savings, which fed into the licence decision
- Wrote a 60-page onboarding guide adopted by teams in India and the Philippines
Research Assistant, Federated Learning
Led a three-person team building HabitatFL, a federated framework for classifying wildlife audio and images across separate clients, with Flower and Docker on multi-GPU infrastructure.
- Trained lightweight edge models (MobileNetV3, SqueezeNet, under 2 MB): 87.65% accuracy on 18.2K spectrograms (27 classes) and 69.92% on 50K images (82 classes) (code)
Research Assistant, Secure Systems and AI
Research on secure messaging and deception-based defence, leading to two peer-reviewed papers.
- Designed and implemented JIGS, a secure-messaging pipeline combining ECDH key exchange with nonce-based AES-128 and parallelised stages, with its protocol verified in ProVerif (code)
- Contributed to Siren, a deception-based defence pairing honeypots and simulated user activity with ML classification of attacker behaviour
Data Science and ML Extern
Built a real-time weapon-detection system on live video with YOLO and Flask, in a team of three.
Education
Bachelor of Technology in Computer Science and Engineering
GPA: 8.69/10
Research
Research Focus
Cross-subject biosignal decoding
Decoding non-invasive biosignals (EEG and wrist surface EMG) so that models trained on some people work on new ones with little or no calibration.
Current Work: Independent entry to the EEG/EMG Foundation Challenge, NeurIPS 2026 (Track 2: EEG decoding): a covariance-based EEG decoder (Riemannian tangent space with logistic regression), written without AI-generated code and evaluated on held-out participants.
Federated personalisation
Personalising shared models without pooling each person's data, using federated learning.
Leakage-free evaluation
Evaluation with subject-wise splits that rules out data leakage.
Research Initiatives
GreyLattice
Founder and writer since February 2025. An independent publication on brain–computer interfaces and computational neuroscience: around ten posts, including long technical deep dives, e.g. a breakdown of real-time MEG decoding of visual perception (arXiv:2310.19812) and an explainer on the limits of non-invasive brain recording. Built and run the publication's own site (Astro, TypeScript, Keystatic CMS).
Learn MoreProjects
Enterprise Banking System (Full-Stack)
Comprehensive banking system with REST API, modern frontend, and containerized deployment.
- Developing REST API with FastAPI and SQLAlchemy ORM for role-based access and ACID-safe transactions
- Building React + TypeScript frontend with Material UI, JWT authentication, and Axios interceptors
- Containerized with Docker Compose, CI/CD via GitHub Actions, 85% test coverage with pytest
GPU-Accelerated Financial News Sentiment
Streamed news through Kafka and scored it with FinBERT, using RAPIDS cuDF on an RTX 2060 to cut end-to-end latency from about 8 minutes to 10 seconds.
LSTM Time-Series Forecasting for Mutual Funds
Advanced ML pipeline for financial forecasting with comprehensive feature engineering.
- Built ML pipeline integrating live NAV API data with S&P BSE Tech index for next-day forecasting
- Engineered 15+ financial features (RSI, MACD, Bollinger Bands, CAPM metrics) achieving MAE 0.3-0.8
- Applied sliding window preprocessing and feature selection using mutual information scoring
JIGSaw - Modular Parallelised Secure Messaging Architecture
High-performance cryptographic communication framework with parallelized execution.
- Designed and implemented parallelized high-performance cryptographic communication system achieving 96.67% speedup
- Developed modular Python architecture with ECDH/AES-128 pipeline and TCP socket implementation
YOLO Weapon Detection System
Real-time computer vision system for security applications.
- Built real-time weapon detection system using YOLO + Flask with optimized TensorFlow inference
- Deployed web application for real-time webcam-based weapon classification
Publications
JIGS: a modular parallel encryption and authentication pipeline for secure messaging
A secure-messaging pipeline combining ECDH key exchange with nonce-based AES-128 and parallelised stages, with its protocol verified in ProVerif.
Siren: Advancing cybersecurity through deception and adaptive analysis
A deception-based defence pairing honeypots and simulated user activity with ML classification of attacker behaviour. Preprint: arXiv:2406.06225.
Certifications
Microsoft Certified: Azure AI Fundamentals
Google Career Certificates: Advanced Data Analytics
Google Cloud Digital Leader
Leadership & Activities
Talk: Introduction to CUDA Programming
Introduction to CUDA programming for 50+ attendees.
Social Media Manager
Strategically managed social media platforms, increasing engagement and visibility for the Startup Incubation Centre