Rudransh Shekhar

Rudransh Shekhar

Founder @ KaptaanLM Research Labs Founder @ mySphere
Former AI/ML Intern @ IMD

B.Tech Computer Science student at VIT Chennai, Founder at KaptaanLM Research Labs & mySphere, and Former AI/ML Intern at the India Meteorological Department (IMD).

My work spans foundation language models, distributed training, cloud infrastructure, and systems engineering. I am building KaptaanLM Research Labs (kaptaanlm.tech) as a platform to conduct, develop, and openly document my independent AI/ML research, focusing on the architecture and distributed pre-training of high-parameter-density small language models (the KaptaanLM family and the Qwen3.5-adapted QaptaanLM-0.75B), large-scale dataset engineering (KapCode-1B, KapInstruct-100M), and local LLM operating systems (InferenceOS).

In parallel, through mySphere (mysphere.co.in), I architect and ship scalable distributed cloud infrastructure, autonomous agentic platforms, and developer tooling. Key systems include myCloud (a distributed personal cloud decoupling metadata from storage nodes with CLI, SDK, and multi-platform clients), myTrade (an autonomous multi-agent market intelligence platform powered by a 40+ LLM consensus jury), developer subdomain & DNS management (myDomain), real-time WebRTC mesh sync (mySync), and computer vision streaming systems (myCam).

At IMD, I worked on AI-driven weather forecasting systems, contributing to the fine-tuning, distributed scaling, and optimization of deep learning models trained on large-scale atmospheric datasets, alongside developing Agentic AI weather assistants with RAG.

Beyond research and venture development, I actively build open-source machine learning systems, multimodal sensor fusion interfaces (Sign Language Interpreter), and high-performance computing pipelines. My goal is to bridge the gap between frontier AI research and reliable, production-grade applications.

Work Experience

Aug 2026 - Present

KaptaanLM Research Labs

Independent AI Research Lab · Foundation Model Development | Available at: kaptaanlm.tech

Founder & Independent AI Researcher | Building & Research | Remote | Aug 2026 - Present
  • Building and steering KaptaanLM Research Labs as an open personal research lab to conduct, engineer, and openly document independent AI/ML research on compact neural architectures, high-parameter-density foundation models, and open pre-training datasets (kaptaanlm.tech).
  • Direct end-to-end foundation model research, distributed pre-training pipelines on Google TPU v5e-8 and NVIDIA GPU clusters, synthetic and domain-specific dataset curation, alignment engineering, and edge quantization runtimes.
  • KaptaanLM (10M – 100M Fleet → Post-100M MoE): Architected an innovative family of dense and sparse small language models (Scout 10M, Vanguard 25M, Navigator 50M, Commander 75M, Sovereign 100M, transitioning to Corsair/Admiral MoE). Pre-trained the baseline Scout-10M model across 5.0 Billion tokens on Google TPU pods using custom 16k BPE tokenization, tied embeddings, QK-Norm, and dual logit soft-capping to resolve small-model repetition collapse.
  • QaptaanLM-0.75B: Adapted Alibaba's Qwen3.5-0.8B-Base by stripping the vision transformer to a lean 752M text causal LM, authoring custom modeling scripts, and executing a two-stage training lifecycle (1B tokens CPT on KapCode-1B and 100M tokens SFT on KapInstruct-100M) on Google TPU v5e-8.
  • InferenceOS: Created an open-source, hardware-aware adaptive operating system for local LLM inference featuring a universal multi-format engine (GGUF, ONNX, PyTorch, OpenVINO), Kahneman System 1 & 2 hybrid symbiosis, dynamic microbatching, live hysteresis layer migration, dual OpenAI/Ollama REST API servers, and an interactive TUI suite.
  • KapCode-1B & KapInstruct-100M: Curated, processed, and published a 1.1-billion-token open-source dataset suite in Apache Arrow and Parquet formats, implementing 50% Fill-in-the-Middle (FIM) infilling and token-level assistant-only loss masking (labels = -100).
  • Multi-Format Open-Source Ecosystem: Authored custom Hugging Face model definitions (modeling_qaptaan.py), published across 8 dedicated repositories, and provided turnkey edge execution via GGUF (13 quantizations with Ollama Modelfiles), BitsAndBytes 4-bit NF4, and in-browser ONNX WebGPU runtimes.
Aug 2025 - Present

mySphere

Udyam Registered MSME Enterprise · Startup | Available at: mysphere.co.in

Founder & Lead Developer | Full-time | Remote | Aug 2025 - Present
  • Founded and lead mySphere, a Udyam registered MSME Enterprise software startup focused on cloud infrastructure, artificial intelligence, distributed systems, and cross-platform product ecosystems (mysphere.co.in).
  • Own product architecture, backend engineering, infrastructure management, deployment pipelines, and end-to-end software development across all mySphere products.
  • myCloud: Architected and built a high-performance distributed personal cloud ecosystem decoupling file metadata from storage nodes with live WebSockets relay, CLI (mycloudctl), Python SDK (mycloud-sdk), Web PWA, Desktop Node GUI, and OCI cloud infrastructure.
  • myTrade: Engineered an autonomous multi-agent AI market intelligence platform for Indian equity markets (NSE/BSE), combining quantitative indicator analysis, 40+ LLM jury consensus, local ChromaDB vector RAG, friction tax modeling, and remote Telegram HITL execution.
  • myCam: Developing a real-time smart IP security surveillance platform with React 19, Socket.IO, WebRTC, Express 5, MySQL, and computer vision motion detection with interactive Canvas activity zones, scheduled recording, two-way audio talk-back, and FIFO loop recycling.
  • myDomain: Developing a developer-focused subdomain registration and dynamic DNS management platform with Cloudflare DNS API integration, multi-record provisioning (A, AAAA, CNAME, TXT, MX), DDNS updates, and 1-click hosting presets.
  • mySync: Building a real-time cross-device clipboard & notification relay platform with WebRTC P2P mesh data channels and Llama 3 AI summary integration across React Native, React + Vite, and Node.js.
  • myTune: Architected a cosmic glass dual-platform music streaming player and desktop application with Electron 33, Vite, Web Audio API multi-band EQ, synchronized lyrics, SQLite library management, and YouTube stream extraction.
  • Designed reusable platform components, shared authentication services (OAuth/2FA), and microservices to optimize scalability, maintainability, and interoperability across the mySphere product suite.
Aug 2025 - Present

India Meteorological Department (IMD)

Ministry of Earth Sciences, Government of India

AI/ML Intern | Internship | Mausam Bhawan, New Delhi, India | May 2026 - June 2026
  • Fine-tuned Huawei's Pangu-Weather model on approximately 3.5 years of IMD GFS data (2023-2026), adapting a globally pre-trained model to regional meteorological forecasting requirements.
  • Optimized code, data-loading, and distributed training workflows, reducing Pangu-Weather training time from about 20 minutes to about 8 minutes per epoch.
  • Contributed to Pangu-Lite, an internally implemented lightweight weather forecasting model, resolving critical training issues and implementing Distributed Data Parallel (DDP) training on over 40 years of ERA5 reanalysis data (1979-2024).
  • Reduced Pangu-Lite training time from about 32 hours to about 5 hours per epoch through distributed computing and pipeline optimization.
  • Built components for an Agentic AI weather assistant using Retrieval-Augmented Generation (RAG), meteorological data sources, and local/cloud-hosted LLMs for natural-language weather forecasting and query processing.
  • Implemented location-aware weather retrieval pipelines using IMD forecasting systems and Mausamgram datasets, integrating Zephyr and Gemma models for contextual responses.
  • Developed PBS job scripts and DDP training workflows using NCCL for multi-GPU and multi-node training across IMD's Arunika High Performance Computing cluster.
  • Refactored fragmented training codebases into production-grade workflows for training, inference, validation, forecasting, checkpointing, early stopping, experiment tracking, geospatial visualization, and model evaluation.
  • Worked with Linux (Red Hat), HPC environments, SSH-based infrastructure, PBS scheduling systems (qsub, qstat), and research-oriented machine learning workflows.
AI/ML Research Collaborator | Part-time | Remote | August 2025 - April 2026
  • Collaborated on AI-driven weather forecasting research using the Pangu-Weather model.
  • Reverse engineered and adapted the open-source Pangu-Weather implementation, fixing model conversion workflows to convert ONNX models into PyTorch (.pth) checkpoints.
  • Validated converted models by comparing ONNX and PyTorch inference outputs, achieving near-identical predictions with negligible numerical error.
  • Developed and optimized auto-regressive forecasting pipelines for multi-step weather prediction, generating forecasts up to 10 days ahead.
  • Worked with GRIB, GRIB2, and NetCDF (NC) meteorological data formats, building scripts for data conversion, preprocessing, validation, and workflow automation.
  • Supported model inference and experimentation across Google Colab TPU infrastructure and institutional HPC systems using Linux, SSH, PuTTY, and command-line workflows.
  • Contributed to optimization, troubleshooting, and validation of large-scale atmospheric data pipelines, model inference workflows, and high-performance computing experiments.

Projects

AI/ML Projects

KaptaanLM KaptaanLM Research Labs Aug 2026 - Present Active Foundation Training · Scout Baseline (5.0B Tok) & Scaling Fleet

High-Parameter-Density Conversational & Reasoning Small Language Models (10M – 100M Dense Fleet → Post-100M Sparse MoE)

PyTorch 2.1+ JAX / Flax 10M–100M → Sparse MoE 16k BPE & Tied Weights QK-Norm & Soft-Capping Hybrid GQA + SWA Google TPU v5e-8 & RTX 6000 Reasoning SFT & CoT
10M–100M → MoEDense Fleet to Sparse MoE
5.0B TokensScout Baseline Pre-Training
58%–90%Weight in Transformer Layers

Project Overview

A family of high-density small language models engineered from scratch for multi-turn dialogue and reasoning. Spanning dense architectures (10M to 100M) and sparse MoE fleets (Armada 1.2B), baseline Scout-10M completed 5.0B token pre-training on Google TPU v5e with steady loss convergence.

Architectural Innovations & Engineering

  • Tied Embeddings + 16k BPE: Custom 16,384 BPE vocabulary paired with tied weights (W_embed = W_lm_head^T) halves embedding footprint, preserving 57.9% (at 10M) up to 90.1% (at 100M) of parameters purely for Transformer reasoning and self-attention layers.
  • QK-Norm (Head-Level RMSNorm): Normalizes query and key head vectors (Q_norm = RMSNorm(Q), K_norm = RMSNorm(K)) prior to scaled dot-product attention, bounding cross-head logits and stabilizing bfloat16 training dynamics under high learning rates.
  • Dual Logit Soft-Capping (Gemma 2 Style): Caps attention scores at 50.0 and output projection logits at 30.0 via tanh scaling, permanently extinguishing degenerative repetition loops, overconfidence spikes, and mode collapse.
  • Hybrid GQA + 512-Token SWA: Combines Grouped-Query Attention with alternating 512-token local sliding window layers across a fixed 2,048-token context window, minimizing KV-cache memory bandwidth for real-time edge generation.
  • Multi-Engine Distributed Training: High-MFU JAX/Flax implementation optimized for Google TPU v5e-8 pods (PJRT runtime, async Orbax checkpointing) alongside native PyTorch + SDPA on NVIDIA RTX 6000 Ada (48 GB VRAM).
  • Curriculum & μP Scaling: Warmup-Stable-Decay (WSD) scheduler with late-stage math and code cool-down injection, utilizing Maximal Update Parametrization (μP) width-scaling to transfer optimal hyper-parameters zero-shot across tiers.
  • Reasoning SFT Ablation Framework: Systematic post-training alignment protocol featuring head-to-head ablation between Branch A (Instruct → Reasoning SFT) and Branch B (Direct Base → Reasoning SFT with <|thought|> scratchpads) on GSM8K, MATH-500, and ARC.
  • Production Verification & Edge Export: 15/15 unit test suite passing (run_tests.py) with turnkey GGUF export (FP16 / Q8_0 / Q4_K_M) for edge acceleration via the InferenceOS Vulkan runtime.

Links & Resources

QaptaanLM-0.75B & Dataset Ecosystem KaptaanLM Research Labs Aug 2026 - Present Stage 1 CPT & Stage 2 SFT Completed · Published

Qwen3.5-0.8B Stripped & Adapted (752M Text-Only) · Two-Stage CPT & SFT · Curated 1.1B Token Ecosystem

Base: Qwen3.5-0.8B 752M Text-Only CPT & SFT Fine-Tuning Google TPU v5e-8 PyTorch / JAX GGUF / BnB / ONNX
752MDense Parameters (ViT Stripped)
1.1B Tokens1.0B CPT + 100M SFT Curated
8 Hub ReposSafetensors, GGUF, BnB, ONNX

Project Overview

A lightweight code generation and reasoning model adapted from Alibaba\'s Qwen3.5-0.8B. Stripped down to a lean 752M text-only causal LM, it was trained on Google TPU v5e across 1.0B tokens of KapCode-1B and 100M tokens of KapInstruct-100M, packaged with turnkey GGUF, CUDA, and ONNX runtimes.

Model Adaptation, Training & Dataset Engineering

  • Architecture Stripping & Adaptation: Adapted from Qwen/Qwen3.5-0.8B-Base by stripping the vision encoder to isolate a 752M text-only causal model with tied embeddings, preserving the 24-layer hybrid linear attention backbone (18 Gated DeltaNet + 6 GQA layers, 256K native context).
  • Custom HF Modeling & Recurrent Caching: Authored custom configuration_qaptaan.py and modeling_qaptaan.py with fast $O(1)$ single-token recurrent caching for DeltaNet linear attention and PyTorch Safetensors export.
  • Stage 1 CPT (KapCode-1B, 1.0B Tokens): Full-parameter continued pre-training on Google TPU v5e-8 across 13 programming languages, docs, and STEM math with 50% Fill-in-the-Middle (PSM FIM), converging loss from 12.80 → 2.44 over 61k steps.
  • Stage 2 SFT (KapInstruct-100M, 100M Tokens): Multi-task instruction tuning across 12 upstream sources in Qwen ChatML with token-level assistant-only loss masking (labels = -100).
  • Turnkey 8-Repository Suite: Distributed across 8 Hugging Face repositories spanning BF16 Safetensors, 13 GGUF quantization tiers (with Ollama Modelfiles), BitsAndBytes 4-bit NF4 / 8-bit Int8 CUDA, and in-browser ONNX WebGPU runtimes.

Links & Resources

Model Hub Repositories:

Dataset Hub Repositories:

Pangu-Weather Fine-Tuning IMD Aug 2025 - May 2026 Completed

Regional Weather Forecasting

Deep Learning PyTorch Weather Modelling Machine Learning Model Inference Distributed Training
20 mins -> 8 mins Reduced Training Time
3.5 YearsTrained on IMD GFS Data
10 DaysAutoregressive Forecast

Project Overview

Fine-tuned Huawei\'s 3D Earth-Specific Transformer on ~3.5 years of IMD GFS atmospheric data (2023–2026) for regional forecasting over the Indian subcontinent. Built end-to-end preprocessing, 10-day autoregressive rollouts, distributed GPU acceleration, and GIS station validation.

Key Engineering & Methodologies

  • Regional Model Adaptation: Fine-tuned Huawei's 3D Earth-Specific Transformer architecture on 3.5 years of IMD GFS data, tailoring pre-trained global representations to Indian regional meteorological dynamics.
  • Large-Scale Data Preprocessing: Engineered automated GRIB-to-NetCDF conversion, humidity parameter derivations, variable normalization, and spatial quality-assurance pipelines for multi-terabyte atmospheric grids.
  • Training & Inference Optimization: Accelerated epoch training time from 20 minutes down to 8 minutes through optimized PyTorch data loaders, mixed precision, and distributed GPU training optimizations.
  • Autoregressive Multi-Day Rollout: Built recursive forecasting pipelines that feed predictions back as sequential inputs to produce stable 10-day autoregressive forecasts across 13 atmospheric pressure levels.
  • Model Verification & Evaluation: Benchmarked predictions against observed ground-station weather data and operational numerical forecast baselines across geopotential height, temperature, humidity, and wind vectors.
  • Geospatial GIS Visualization: Developed automated visualization pipelines mapping meteorological forecast fields to GIS coordinate systems for operational forecasting analysis.

Links & Resources

Pangu-Lite IMDMay 2026 - June 2026 Completed

Distributed AI Weather Forecasting Model

Distributed Computing Deep Learning HPC NCCL Model Training ERA5 Data Forecasting
32 hours -> 5 hours Reduced Training Time
40+ YearsTrained on ERA5 Data
Multi-GPU Multi-Nodal DDP Setup

Project Overview

A lightweight deep learning weather forecasting model trained on 40+ years of ECMWF ERA5 reanalysis data (1979–2024). Engineered a unified production pipeline, implemented multi-GPU multi-node DDP training on IMD\'s Arunika HPC cluster, and explored patch embeddings with cross-attention.

Key Innovations & Distributed Training

  • Lightweight Deep Learning Model: Co-developed a lightweight atmospheric forecasting model trained across 40+ years of ERA5 reanalysis grids, delivering fast regional weather predictions with reduced compute footprint.
  • Codebase Refactoring & Pipeline Unification: Resolved critical legacy training bugs and refactored disparate scripts into modular, production-ready training, validation, checkpointing, and logging pipelines.
  • Multi-Node DDP on Arunika HPC: Engineered PyTorch Distributed Data Parallel (DDP) workflows with NCCL backends and authored custom PBS job scripts for scalable multi-GPU, multi-node execution on IMD's Arunika cluster.
  • 6.4x Training Acceleration: Reduced epoch training duration from 32 hours to 5 hours per epoch through optimized NCCL distributed communication, pinned-memory data loading, and kernel execution tuning.
  • End-to-End MLOps Workflows: Built resilient checkpointing, automated experiment logging, loss monitoring, and autoregressive multi-step inference pipelines for systematic model iteration.
  • Architectural Experimentation: Researched and benchmarked 3D patch embedding variants, variable-level cross-attention mechanisms, ensemble inference strategies, and post-hoc error correction.

Links & Resources

InferenceOS KaptaanLM Research Labs July 2026 - Present Active Development

Hardware-Agnostic Adaptive Operating System for Local LLM Inference

Python 3.10+ C++20 CUDA & ROCm Vulkan & Metal Multi-Format Engine System 1 & 2 Symbiosis FastAPI (OpenAI & Ollama) Modern TUI & CLI
System 1 & 2Kahneman Hybrid Symbiosis
Multi-FormatGGUF, ONNX, PyTorch, OpenVINO
Zero OOMDynamic Microbatch & Migration

Project Overview

An open-source, hardware-aware adaptive OS and control plane for local LLM inference. Forked from llama.cpp with a multi-format engine, it features a <150ms System 1 reflex engine, System 1+2 hybrid gating, dynamic VRAM layer placement, KV cache compression, dual REST APIs, and a modern TUI.

Features

  • System 1 Non-Autoregressive Reflex Engine: Sub-150ms instant encoder classification (Laya mmBERT, Kev pointer heads) with zero KV cache footprint for sub-second decision making.
  • Kahneman Fast/Slow Hybrid Symbiosis: Dynamic cognitive loop combining sub-second reflexes with generative deep-reasoning LLMs, calibrated confidence gating (τ), automatic System 2 escalation, and DAgger distillation logging.
  • Universal Multi-Format Execution: Polymorphic engine with magic-byte format autodetection supporting GGUF, SafeTensors, ONNX, PyTorch, Intel OpenVINO, Keras 3, TFLite, and dynamic INT8 quantization.
  • Hardware Auto-Profiler & Cost Optimizer: Multi-vendor detection engine profiling CPU, NVIDIA CUDA, AMD ROCm, Apple Metal, Intel Arc, and integrated GPUs to auto-generate optimal execution and placement plans.
  • Dynamic Microbatch Scheduler & Zero OOM: Real-time prefill batching using multi-heuristic scoring (VRAM headroom, prompt TPS, GPU occupancy, PCIe bandwidth) to eliminate out-of-memory crashes.
  • Live Hysteresis Layer Migration: Dynamic offload coordinator that shifts layers in real-time between dGPU, iGPU, and CPU under memory pressure during active inference.
  • KV Cache Compression & Eviction: Memory optimizer supporting FP16/Q8_0/Q4_0 cache quantization and attention-sink eviction policies (H2O Heavy-Hitter Oracle, StreamingLLM, LRU, FIFO).
  • Modern Interactive TUI & Unified CLI: Terminal UI featuring floating slash autocompletion (/model, /mode, /tau), modal split-view model picker (Ctrl+O), persistent status ribbon, and rich deliberation flowcards across a 27-command CLI suite.
  • Dual REST API Server: High-throughput FastAPI HTTP server supporting full OpenAI (/v1/*), Ollama (/api/*), and System 1 (/v1/system1/*) endpoints with streaming SSE.

Links & Resources

myTrade mySphere May 2026 - Present Internal Testing

Autonomous AI Market Intelligence & Trading Platform

Multi-LLM Consensus ChromaDB (RAG) SQLite Upstox API Streamlit UI Telegram Bot AsyncIO
Multi-LLM Jury40+ Models Consensus
Quant EngineMath Risk Veto
HITL & AutoTelegram Control

Project Overview

An autonomous, multi-agent AI market intelligence and trading system for Indian equity markets (NSE/BSE). Combines quantitative indicators, a 40+ LLM jury consensus, local vector RAG memory, and strict risk guardrails with a remote Human-in-the-Loop (HITL) control interface.

Features

  • Multi-Model Consensus Engine: Queries a configurable pool of independent LLMs concurrently (40+ model roster across Groq, Nvidia, Mistral, Google, Cerebras) with dynamic multi-key API cycling to benchmark qualitative sentiment without single-model hallucinations.
  • Mathematical Risk Veto Layer: Deterministic quantitative rules evaluating macro market regimes, price action, and momentum to serve as a pre-AI guardrail layer that can veto qualitative entries.
  • Hybrid Data & RAG Memory System: Combines SQLite structured price storage with ChromaDB vector embeddings for corporate disclosures, regulatory filings, and market news.
  • Google OKF Context Assembly: Synthesizes price metrics, technical indicators, and vector RAG context into standardized Open Knowledge Format (OKF) payloads for model evaluation.
  • Friction & Tax Modeling Engine: Calculates transaction friction costs including STT, GST, SEBI charges, brokerage, and stamp duty to verify net capital deployment efficiency.
  • Interactive HITL & Telegram Control: Secure 4-message Telegram dashboard allowing real-time monitoring of AI juror votes, news context, and tax breakdowns with interactive trade approval, rejection, or veto override buttons.
  • Execution & Broker Integration: Features a local SQLite paper ledger for shadow portfolio tracking alongside Upstox API integration and configurable auto-approval switches.

Links & Resources

Agentic AI Weather Assistant IMD May 2026 - June 2026 Completed

LLM-Powered Meteorological Query System

Retrieval-Augmented Generation (RAG) Large Language Models (LLM) Natural Language Processing (NLP)
RAGArchitecture
Zephyr/GemmaLocal & Cloud LLMs
MausamgramWeather Data

Project Overview

An Agentic RAG meteorological assistant developed at IMD to translate numerical weather forecasts into accessible insights. Synthesizes observation feeds and Mausamgram data with local (Zephyr) and cloud (Gemma) models for location-aware querying and automated alert generation.

Key Architecture & Features

  • Agentic RAG Meteorological Pipeline: Engineered an intelligent conversational agent that orchestrates multi-step information retrieval across real-time IMD numerical forecast systems and structured meteorological tables.
  • Location-Aware Data Retrieval: Developed ingestion pipelines parsing localized Mausamgram observation datasets to answer station-specific temperature, precipitation, and storm alert queries.
  • Hybrid Local & Cloud LLM Inference: Implemented dual inference runtimes integrating local quantized Zephyr models for air-gapped on-premise inference and cloud-hosted Gemma models via Hugging Face endpoints.
  • Context Engineering & Query Synthesis: Built automated context-assembly pipelines that normalize numerical weather tables into formatted LLM prompts for hallucination-resistant domain responses.
  • Pipeline Hardening & Optimization: Refactored and optimized the core RAG retrieval flow, mitigating latency bottlenecks and significantly improving response reliability during peak forecast query loads.
  • Conversational Forecast Generation: Enabled conversational explanation of severe weather bulletins, regional climate trends, and daily forecasts in clear, actionable natural language for public advisories.

Links & Resources

Development & Product Projects

myCloud mySphere Aug 2025 - March 2026 Live

Distributed Personal Cloud Platform

Flask MySQL Socket.IO Python SDK CLI (mycloudctl) CustomTkinter Oracle OCI Nginx & Gunicorn
Cross-PlatformWeb, Desktop, Android, CLI, Python-SDK
Real-TimeSocket.IO Sync
DistributedStorage Nodes

Project Overview

A full-stack distributed personal cloud platform engineered for the mySphere ecosystem. Decouples file metadata from physical storage, enabling personal devices to act as distributed "Smart Relay" nodes with centralized control, real-time sync, enterprise encryption, and CLI/SDK/GUI access.

Features

  • Multi-Client Access: Responsive Web PWA, CustomTkinter Desktop Node GUI, Python CLI (mycloudctl), official Python SDK (mycloud-sdk), and cross-platform Electron/Capacitor wrappers.
  • Distributed Storage Nodes: Connect private PCs or servers as storage nodes over WebSockets ("Smart Relay") with MySQL node routing.
  • Advanced Sharing Suite: Direct user invites with RBAC permissions, public token link generation, and batch share links grouping disparate files/folders.
  • Stash Lifecycle: Temporary file isolation workflow with scheduled auto-cleanup (10m, 1h, 1d, or custom intervals).
  • Authentication & Security: Google/GitHub OAuth, 2FA challenge flows, CSRF protection, HttpOnly/SameSite session security, and bcrypt hashing.
  • Observability & Monitoring: Built-in internal dashboards monitoring system CPU/RAM, Nginx/Gunicorn service health daemons, and product analytics.
  • Infrastructure & Hosting: Production-grade deployment on Oracle Cloud Infrastructure (OCI) using Nginx reverse proxy and Gunicorn Eventlet async worker processes.

Links & Resources

myCam mySphere VIT/College In Development · Lead Architecture & Core Systems

Real-Time WebRTC Security & Surveillance Platform

React 19 Vite 7 Express 5 Socket.IO WebRTC MySQL Tailwind CSS Canvas API AI Motion Detection Two-Way Audio
WebRTC P2PLow-Latency Stream
AI DetectionPerson/Vehicle/Pet
Two-Way AudioFull Duplex Talk-Back

Project Overview

A full-stack real-time video surveillance platform built with React 19, Express 5, WebSockets, and WebRTC. Turns connected devices into P2P security feeds with AI object detection, custom canvas activity zones, full-duplex two-way audio, scheduled recording, and automated loop recycling.

Features

  • Turnkey Phone-as-CCTV Broadcasting: Transforms smartphones, laptops, or webcams into live broadcasting security cameras with low-latency WebRTC P2P streaming and Google STUN NAT traversal.
  • AI Object Detection & Zone Filtering: Computer vision engine classifying target entities (Person, Vehicle, Pet) with Ray-Casting Point-in-Polygon intersection checks, dynamic bounding boxes, and automated alert snapshot triggers.
  • Interactive Canvas Activity Zones: HTML5 Canvas overlay for drawing custom polygonal detection zones, saved as normalized coordinates to MySQL to filter false alarms.
  • Full-Duplex Two-Way Audio Talk-Back: Bidirectional WebRTC audio channel enabling seamless real-time talk-back from the viewer dashboard to camera speakers.
  • Automated Scheduled Recording: Rule manager supporting time windows (HH:MM), day-of-week recurrence, and background evaluation loops for unattended scheduled recording.
  • Dynamic Bitrates & Multi-Tier Storage: Quality switching (4K UHD 8Mbps, 1080p HD 3Mbps, SD 800kbps) with auto-failover storage (Internal → SD Card → Cloud) and FIFO loop recycling that purges oldest clips when quota caps are reached.
  • Interval Auto-Snapshots & Smart Alerts: Configurable automated snapshot intervals (15s, 30s, 60s), real-time Socket.IO alerts, in-app banners, and native browser Web Push desktop notifications.
  • myCloud Ecosystem Integration: Shared authentication with scrypt/pbkdf2 password hashing and direct video clip uploads to myCloud personal storage.

Links & Resources

myDomain mySphere Oct 2026 - Present In Development

Developer Subdomain & Cloudflare DNS Management Platform

React 18 TypeScript Vite Express 5 Cloudflare DNS API Supabase / PostgreSQL Dynamic DNS (DDNS) Tailwind CSS OAuth 2.0
Cloudflare APIAutomated DNS Sync
Multi-RecordA, AAAA, CNAME, TXT, MX
DDNS & Presets1-Click Vercel / GitHub Pages

Project Overview

A full-stack developer platform providing free custom subdomains under configured parent domains with real-time DNS management. Built with React, Vite, TypeScript, Express 5, and Cloudflare, featuring Dynamic DNS (DDNS) automation for homelabs, 1-click deployment presets, and global propagation diagnostics.

Features

  • Developer Subdomain Provisioning: Instant self-service registration and lifecycle management of subdomains under parent domains with reserved keyword filtering and abuse prevention.
  • Full-Spectrum DNS Record Control: Real-time management of standard DNS records (A, AAAA, CNAME, TXT, MX, NS, SRV, CAA) and 301/302 URL redirects.
  • 1-Click Hosting Integration Presets: Pre-built DNS templates for Vercel, GitHub Pages, Netlify, Render, and Cloudflare Pages for friction-free developer deployments.
  • Dynamic DNS (DDNS) Update Endpoint: Dedicated HTTP token endpoints enabling self-hosters and home-lab servers on dynamic public IPs to synchronize DNS records automatically.
  • Cloudflare DNS Integration & Proxying: Bidirectional synchronization with Cloudflare API, automated zone updates, configurable TTLs, and instant proxy toggle (orange-cloud).
  • Global Propagation Inspector: Diagnostic resolver suite querying worldwide authoritative DNS servers in real time to verify propagation status and health.

Links & Resources

mySync mySphere In Development

Cross-Device Notification & Clipboard Sync Platform

Node.js / Express Socket.IO WebRTC React Native React + Vite Groq / Llama 3 AI Clipboard Sync

myTune mySphere Jan 2026 - Present Completed · Dual-Platform Architecture

Cosmic Glass Cross-Platform & Desktop Music Streaming Application

Electron 33 JavaScript / Node.js Vite Web Audio API SQLite / sql.js Motion music-metadata Custom Streaming Protocol Cross-Platform

Integrated Systems Projects

Sign Language Interpreter VIT/College Aug 2026 - Present Live Production Web Suite & Active Research

Sign Language Studio · Real-Time ASL Vision Suite, Text-to-Sign Synthesis, Windows OS Automation & WebRTC

MediaPipe Hands TensorFlow.js WebRTC (PeerJS) Windows OS Automation WebSocket Bridge GPT-4o Vision Assistant ESP32 & BLE (Optional) 14-Language Speech Synthesis Netlify Production
Web ProductionSign Language Studio Suite
Neural VisionMediaPipe & On-Device DNN
OS AutomationTouchless Windows Control

Project Overview

A production-grade American Sign Language (ASL) recognition and assistive communication web suite deployed at sign-language-studio.webdomain.live. Fuses 21-landmark browser computer vision with an interactive practice academy, bidirectional Text & Speech-to-Sign telecommunication, touchless Windows OS gesture automation, and multimodal AI diagnostics.

Features & Technical Highlights

  • Full Production Web Suite (Sign Language Studio): Deployed web platform featuring specialized interconnected studios for real-time camera detection, gamified speed quizzes, multimodal AI chat, two-way telecommunications, and an interactive 46-sign visual dictionary.
  • 21-Landmark Vision Tracking & On-Device DNN: Client-side tracking using Google MediaPipe Hands capturing 21 3D hand keypoints at 30–60 FPS, normalized and classified into the 26 ASL alphabet letters via a 4-layer Sequential DNN cached in IndexedDB.
  • Windows OS Gesture Automation & Companion Bridge: Fully operational companion bridge (python run_os_control.py / standalone executable) over WebSocket (ws://127.0.0.1:8765) for touchless Windows automation: smoothed cursor steering, pinch click/drag, scrolling, volume dial, brightness, snipping tool, and desktop virtual typing injection.
  • Interactive Practice Academy & Speed Quiz: Gamified learning laboratory featuring pose hold-and-verify SVG progress rings, streak multiplier scoring, and a 60-second speed-run evaluation engine.
  • Bidirectional Text & Speech-to-Sign Synthesis: Full-duplex assistive engine converting typed text prompts and spoken voice into animated ASL sequences, paired with WebRTC P2P communication and real-time sign-to-speech translation.
  • Multilingual Vocalization & Predictive Text: Offline Trie and bigram language model for fast sentence autocomplete, coupled with real-time SpeechSynthesis vocalization and cloud translation across 14 languages.
  • Multimodal AI Assistant & Active Learning: Puter.js GPT-4o Vision assistant for analyzing uploaded sign photos, paired with instant k-NN calibration for user-specific hand shapes without model retraining.
  • Hardware Disambiguation & Smart Glove Telemetry: 100% functional with standard webcams alone; includes optional research extension fusing an ESP32 + MPU-6050 BLE glove streaming 6-axis kinematics at 50 Hz for dynamic motion gestures ('J', 'Z', 'YES').

Links & Resources

EMG Controlled Music System VIT/College Oct 2025 - March 2026 Completed

Bio-Signal Interface Project (Arduino)

Arduino Embedded Systems EMG Signals C/C++ DFPlayer Mini Ultrasonic Sensing OLED Interface

Other Projects

Renewable Energy DB Simulator VIT/College Oct 2026 Completed

Relational Database Simulator & Streaming IoT Telemetry Platform

React 19 TypeScript Vite Relational DBMS Interactive ER Canvas Streaming IoT Telemetry Motion Lucide Icons

Open Source & Community Contributions Open Source 2026 Active Contributor

Upstream Foundation ML, Systems Emulation & Developer Tooling Contributions

Python Hugging Face Transformers Audio ML / CLAP C++ / Win32 Winsock API Systems Emulation Node.js CLI Google Antigravity Agent Skills Cross-Platform POSIX Git / GitHub PRs

Shor's Algorithm Implementation VIT/College Feb 2026 - April 2026 Completed

Quantum Computing Project (Qiskit)

Qiskit Quantum Algorithms Python QFT Modular Exponentiation Classical Post-Processing Circuit Optimization Beauregard Architecture RSA Crypto Analysis Web Simulation (HTML/CSS)

Education

Vellore Institute of Technology

Bachelor of Technology - BTech, Computer Science | 2025 - 2029

Grade: CGPA: 9.44

  • Achieved a CGPA of 9.60 in Semester 1 and a cumulative CGPA of 9.44 after completion of the first year.
  • Secured 3rd rank in the branch during Semester 1 and maintained strong academic performance throughout the first year.
  • Coursework: C/C++, Python, Shell/Bash, Linux, SQL, R, MATLAB, Data Structures & Algorithms, Digital Electronics, Computer Architecture, SolidWorks, LTSpice.

Senior Secondary Education (Class 12th) | 2025

R.N. Public School, Prayagraj (CBSE Board)

Grade: 91.6%

  • Secured 91.6% overall in the Class 12th CBSE Board examinations.
  • Demonstrated academic excellence with strong and consistent performance in technical and analytical subjects.
  • Highlights: Scored 95/100 in Computer Science, 98/100 in English and 90/100 in Physics.

Senior Secondary Education (Class 11th) | 2024

R.N. Public School, Prayagraj (CBSE Board)

Grade: 91.1%

  • Secured 91.1% overall in the Class 11th CBSE examinations.
  • Demonstrated strong academic foundation in Computer Science and Mathematics.
  • Highlights: Scored 93/100 in Computer Science and 92/100 in Mathematics.

Secondary Education (Class 10th) | 2023

St. Joseph's College, Prayagraj (ICSE Board)

Grade: 96.8%

  • Secured 96.8% overall in the Class 10th ICSE Board examinations.
  • Achieved a perfect score of 100/100 in Computer Applications.
  • Scored 97/100 in English, 96/100 in Mathematics, 95/100 in Science, and 96/100 in History, Civics and Geography.

Certifications & Courses

Machine Learning for Earth Systems Modelling | ECMWF

Comprehensive training on ML foundations for weather and climate science, neural architectures for atmospheric dynamics, deterministic AI forecasting systems (AIFS), and data-driven forecasting frameworks.

View Foundations Certificate View Architectures & Prediction Certificate

Understanding climate data: From analysis of the past to future projections | ECMWF

Comprehensive training on climate data analysis workflows, ERA5 reanalysis and historical observational baselines, future projection methodology (CMIP6), climate impact indicators, and assessing uncertainty in long-term atmospheric trends.

View Certificate

Fundamentals of Remote Sensing & Air Quality | NASA ARSET

Core principles of remote sensing, sensor types, data resolution, satellite imagery interpretation, and satellite data applications for atmospheric air quality monitoring.

Codes: 4794012384RS | 3979347422RS
View Fundamentals Certificate View Air Quality Certificate

Hyperspectral Data for Land and Coastal Systems | NASA ARSET

Advanced techniques for processing and analyzing imaging spectroscopy datasets, exploring airborne and spaceborne hyperspectral sensors (EMIT, PRISMA, PACE), spectral signature extraction, and monitoring vegetation health, aquatic habitats, and coastal water quality.

Verification Code: 2804106189RS Verify ↗
View Certificate

Advanced Features in Java | Infosys Springboard

Professional certification covering core enterprise Java capabilities, focusing on the Java Collections Framework, working with Sets, stream processing, lambdas, and object-oriented architectural patterns.

Issued: October 2026 Verify ↗
View Certificate

Understanding satellite data to monitor dust | EUMETSAT

Practical application of geostationary and polar satellite imagery (Meteosat SEVIRI Dust RGB, Sentinel, Metop) to detect, track, and forecast airborne dust storms, evaluate aerosol optical depth (AOD), and analyze transboundary atmospheric transport.

View Certificate

H SAF Products (Snow, Soil Moisture, Precipitation) | EUMETSAT

Comprehensive training on H SAF satellite products focusing on snow cover, soil moisture monitoring, and precipitation measurements for meteorological and hydrological applications.

View Snow Certificate View Soil Moisture Certificate View Precipitation Certificate

Honours & Rewards

Certificate of Merit - Branch Rank 3 | 2026

Vellore Institute of Technology (VIT), Chennai

  • Awarded for securing 3rd rank in the CSE branch.
  • Award was presented by Mr.Binod Kumar, MD & CEO of Indian Bank.
  • Award included a certificate.
View Certificate

Academic Excellence Scholarship | 2025

Indian Farmers Fertilizer Cooperative Limited (IFFCO)

  • Awarded by Indian Farmers Fertilizer Cooperative Limited (IFFCO) for securing 91.6% in the CBSE Class 12th Board Examination in 2025.
  • Award included a certificate, medal and a scholarship reward of Rs. 3,28,000

Academic Excellence Scholarship | 2023

Indian Farmers Fertilizer Cooperative Limited (IFFCO)

  • Awarded by Indian Farmers Fertilizer Cooperative Limited (IFFCO) for securing 96.8% in the ICSE Class 10th Board Examination in 2023.
  • Award included a certificate, medal and a scholarship reward of Rs. 24,000
View Certificate

Certificate of Excellence in Computer Applications | 2023

St. Joseph's College, Prayagraj

  • Awarded by St. Joseph's College for scoring a perfect 100/100 in Computer Applications in the ICSE Class 10th Board Examinations in 2023.
  • Award included a certificate and a trophy.
View Certificate

Featured in Amar Ujala for Board Results | 2023 & 2025

Amar Ujala Newspaper

  • Mentioned for securing 96.8% in Class 10th (ICSE).
  • Mentioned again for securing 91.6% in Class 12th (CBSE).
  • Recognized in merit-focused board result coverage for consistent academic achievement.
View Article Mention

Featured in CollegeDekho VITEEE Toppers List | 2025

CollegeDekho.com

  • Recognized in the VITEEE toppers list feature.
  • Included in the published ranking highlights among notable VITEEE performers.
View Feature Highlight

Skills & Technologies

Artificial Intelligence & ML

PyTorch, JAX / Flax, Distributed Training, Model Inference, RAG, LLMs, Agentic AI, ChromaDB, TensorFlow.js, MediaPipe Hands, NLP, CPT, SFT, SLMs

Backend & Systems Engineering

Python, C / C++, Java, Node.js / ExpressJS, Flask, REST APIs, WebSockets (Socket.IO), WebRTC P2P, llama.cpp Engine

Cloud, GPU & Infrastructure

HPC, Distributed Computing (NCCL, DDP), CUDA, ROCm, Vulkan, Metal, Oracle Cloud (OCI), Nginx, Gunicorn, systemd

Data & Storage Systems

Apache Arrow & Parquet, Dataset Curation & Sharding, SHA-256 Deduplication, SQL, MySQL, SQLite, ChromaDB Vector DB, GRIB/NetCDF, Geospatial GIS

Frontend & App Development

React, React Native, Vite, CustomTkinter, HTML, CSS, JavaScript, UI/UX, Electron, Capacitor, PyQt6, Canvas API

Security & Systems Architecture

OAuth 2.0, 2FA, System Design, Live Hysteresis VRAM Migration, Dynamic Microbatching, Distributed Storage

DevOps & Open Source Publishing

Git, GitHub, Hugging Face Hub / Datasets, Kaggle Datasets, CI/CD, Docker, Linux Administration, PyPI Publishing (mycloudctl, mycloud-sdk, inferenceos)

Hardware & Embedded Systems

ESP32, MPU-6050 6-Axis IMU, Web Bluetooth (BLE), Sensor Fusion, Arduino C++, Bio-signals (EMG), Ultrasonic Sensing, DFPlayer Mini, OLED Displays, Serial Protocols

Languages

English (Fluent), Hindi (Native), German (Elementary)

Contact/Socials

Phone : +91 8429632553

GitHub : github.com/rudy-07

Hugging Face : huggingface.co/kaptaan45

Kaggle : kaggle.com/kaptaan45

Website : mysphere.co.in

Research Labs : kaptaanlm.tech

LinkedIn : linkedin.com/in/rudransh-shekhar/

Email : kanhaiffco2007@gmail.com

Resume & Profile

Master Technical

Foundation AI & Research Resume

Typeset in LaTeX with an ATS-calibrated, 2-page format. Comprehensive overview of foundation model pre-training at KaptaanLM Labs, atmospheric deep learning HPC scaling at IMD, distributed mySphere systems, and academic honours.

Foundation Models (SLMs) TPU Pod Pre-training (5B Tokens) Pangu Atmospheric HPC Scaling Agentic RAG & LangModels KapCode-1B Dataset
View In Browser

9.44

B.Tech CGPA

3rd

Branch Rank

5.0B

Pretrained Tokens

1.1B+

Open Dataset

40+

LLM Jury Models

AI/ML

Core Focus

Download PDF
About
Experience
Projects
Education
Certifications
Honours
Skills
Contact
Resume