Shubhankar Mahanta
Deep learning & Machine Intelligence
I'm a machine learning fella currently pursuing my Bachelor's. I work with datasets, training pipelines, model reasoning, GPU-accelerated environments, and Deployment Management. I specialize in PyTorch and quantized transformer pipelines, balancing latency, throughput, and accuracy to deliver high-performance, production-grade intelligence.
Selected Projects
Reward model for SFT
For post training and alignment with human preference
Reward Model for SFT Alignment trains transformer-based scorers to predict human preferences, generating calibrated reward signals for RLHF fine-tuning. It processes paired preference data via Hugging Face tokenizers and PyTorch Datasets, supports GPT-2/BERT backbones with mixed-precision, gradient checkpointing, and optimizes a pairwise logistic loss across distributed NCCL-enabled GPUs.
Audio MEL fingerprinting
Advanced audio processing and feature extraction
AudioEffects_DS provides a high-performance modular pipeline to augment audio datasets by slicing full-length tracks into configurable segments, applying advanced parametric DSP effects, and extracting metadata for downstream ML tasks. Built with Python, torchaudio, and NumPy, it supports randomized reverberation, equalization, compression, pitch shifting, and noise injection.
ML Suite
Context Builder, Relation Graph Computation, CoT Chain-of-Thought, and KOG Knowledge Graph
Context Builder retrieves relevant passages using FAISS vector search. Relation Graph Computation builds multi-hop relational graphs with DGL/PyTorch. The CoT Chain-of-Thought framework orchestrates sequential sub-prompts with intermediate activations logged for structured reasoning. KOG Knowledge Graph ingests diverse sources into RDF triples, normalizes entities with ontology mapping, and offers SPARQL endpoints.
Rust Kyber MQ
Post-quantum secure messaging
Kybr-ME is a self-hosted, end-to-end encrypted messaging service written in Rust and secured by the Kyber768 post-quantum KEM. Leverages Tokio and Actix-Web for asynchronous WebSocket transport, providing both one-to-one and group chats with perfect forward secrecy. Each session begins with a Kyber768 KEM handshake.
uNet-GAN
Image/Video Generation Pipeline
UGen-v01 is an image generative model and pipeline supporting YAML-driven pipeline configs, batch inference, and mixed-precision execution on CPU/GPU. Core components handle device-aware model loading, pipeline chaining, and flexible sampling strategies. Integrated with Weights & Biases for experiment tracking.
