Adaptive TabNet–β-VAE Fusion Ensemble Closes Uncertainty Gap In Fraud Detection
Multi-stage deep learning pipeline for fine-grained emotion analysis and persuasion detection. Custom SGEE layer fuses RoBERTa embeddings with NRC-VAD lexicon features into a 512-dim emotion space. 6 specialized prediction heads across 28 emotion classes, VAD regression, and persuasion tactics.
Lightweight Conformer-Lite architecture for music genre classification with artist-disjoint evaluation. Exposes artist leakage in standard benchmarks (18.8% F1 drop) using gradient reversal for artist-invariant representations. Multi-exit early stopping for edge deployment.
End-to-end emotion detection across text, image, audio, and video via a unified REST API. Custom CNN → BiLSTM → Transformer → Attention model (UltimateNet) for audio. 7-stage video pipeline with Whisper transcription, EasyOCR, and Cerebras LLM cross-modal fusion.
1M synthetic transactions with realistic device/account behavior simulation. TabNet–β-VAE ensemble for anomaly detection. Published as first-author research paper at ICAIN 2025 (Scopus & Springer indexed).
Multi-threaded client-server chat app built from scratch in Java. Custom TCP/IP socket layer, thread-per-client architecture with ConcurrentHashMap, normalized MySQL schema, MVC pattern, and custom Swing UI with message bubbles, presence detection, and read receipts.
Hybrid ResNet-50 + Swin Transformer classifier with a custom Gated Attention mechanism. Trained with AdamW + CosineAnnealingWarmRestarts. Served via Flask REST API with a full web UI — drag-and-drop upload, prediction history, dark mode, and user auth.
YOLOv8 detection + tracking + ROI pipeline for autonomous drone missions. Optimized for edge deployment on NVIDIA Jetson Nano under real-time latency constraints. Core system behind Team Griffin's AIR 1 performance at SAE Aeromodelling Challenge 2025.
15,000+ lines of C across 15+ modules. Full-featured banking platform with multi-session concurrency (Win32 threads), flat-file persistence, live stock market simulation (multi-factor price model), quantitative analytics, 4-heuristic fraud detection engine, 2FA, and an admin panel.
Official production website for a high-growth agritech company — live and serving real users. Full backend: Express.js, Supabase auth + database, role-based access control, and Razorpay payment gateway. Delivered in 4 weeks via MLSC industry project.
What I do when I'm not building things — and why it matters.
Discipline over motivation. The barbell doesn't care about your mood — you either lift it or you don't. Progressive overload taught me compounding before any textbook did.
Forced mindfulness — real-time systems thinking at 80 km/h. You can't be distracted on a bike. Some of my best ideas come mid-ride, and some of my clearest decisions do too.
Technique over brute force. Trying harder doesn't mean going faster — reduce drag before you increase power. That lesson transferred directly to how I approach engineering problems.
Fast, reactive, and deeply strategic. Doubles sharpened my teamwork instincts — anticipate what your teammate needs, cover their gaps, and trust them to cover yours.
Both instruments taught me layered thinking — rhythm underneath melody, structure underneath expression. Same as designing a good system: every component has its role.
Constraints breed creativity. One instrument, one challenge: make it sound like a full production. Same principle that built a 15,000-line banking system in C with no external libraries.
Always open to interesting builds, research, and conversations.