Everything I have engineered, opened up. Full-stack web architectures, machine learning models, and strict perimeter limits.

4 PRODUCTION BUILDS
Full-Stack • Autonomous Systems • RAG AIPRODUCTION PLATFORM
GDACS Live Satellite StreamNDMA & FEMA Protocol-Grounded RAG
Index: 01 / 04Build notes →

Aegis Disaster Preparedness

Autonomous disaster management, live GDACS telemetry, and protocol-grounded RAG AI

A comprehensive institutional disaster response platform built for Smart India Hackathon 2025 (Top 10 Finalist). Integrates real-time geospatial satellite feeds from GDACS, autonomous emergency alert broadcasting, 8 interactive bilingual disaster safety modules, and an NDMA/FEMA protocol-grounded RAG chatbot.

Hard Limit (What it refuses to do):

Never hallucinates medical triage: every emergency response strictly references official NDMA/FEMA protocols

Disaster Modules8 Complete
AI GroundingNDMA/FEMA
Node.jsExpress.jsMongoDB AtlasGemini 2.5 FlashGDACS APIViteTailwind CSSGSAP
Machine Learning • Signal ProcessingRESEARCH CASE STUDY
EEG Signal Analysis · SwaarogymTheta/Beta Ratio · Welch PSD Extraction
Index: 02 / 04Build notes →

EEG Cognitive Attention System

Cognitive attention monitoring and interactive game feedback using theta-beta spectral ratios

Developed during the research internship at Swaarogym Medical Devices (MIIC, MNIT Jaipur). A machine learning system analyzing EEG frequency bands to classify user attention levels during cognitive tasks. Computes Theta/Beta power spectral ratios using Welch's method to drive an interactive feedback game.

Hard Limit (What it refuses to do):

Never claims clinical diagnosis: model is strictly designed for educational cognitive monitoring feedback

Random Forest Acc93.0%
Dataset Samples36 Subjects
PythonSciPy (Welch PSD)Scikit-LearnNumPyMatplotlibPandas
Predictive Analytics • Enterprise ML • MLOpsLIVE WEB APP
Predictive Lead Scoring EngineProduction Scikit-Learn Pipeline
Index: 03 / 04Build notes →

SignalScale (Lead Scoring)

Real-time enterprise lead conversion scoring and predictive analytics engine

A high-throughput predictive machine learning system designed to score inbound B2B sales leads in real time. Features gradient boosting classification, automated feature transformation pipelines, and dynamic conversion probability distribution gauges.

Hard Limit (What it refuses to do):

Refuses uncalibrated raw logit outputs: enforces Platt probability calibration

Inference Latency<15ms
ArchitectureGradient Boost
PythonScikit-LearnStreamlitPandasNumPyJoblibRender Cloud
Deep Learning • Computer Vision • Canvas DSPLIVE WEB APP
Interactive CNN ClassifierCentroid Alignment & Softmax Top-3
Index: 04 / 04Build notes →

Handwritten Digit Recognition

Real-time interactive canvas drawing and Convolutional Neural Network inference

An interactive web-based deep learning application that performs real-time classification of hand-drawn digits. Features custom client-side bounding box normalization, center-of-mass centering, and top-3 softmax confidence distribution visualization.

Hard Limit (What it refuses to do):

Refuses blind resizing: enforces center-of-mass translation to preserve digit morphology

Test Accuracy99.2%
Downsampling28x28 Centroid
PythonPyTorch / CNNHTML5 CanvasFlask / RESTNumPyCSS3