// projects

Things I've Built

// featured & award-winning

Real-Time Electricity Demand Forecasting

SIH'24 Grand Finale Winner
AI / MLSep 2024 — Dec 2024

Problem: Delhi's power grid lacked accurate short-term demand forecasting. Solution: Built SARIMAX + Transformer hybrid model on 12+ years of historical data enriched with real-time weather. Impact: 98%+ accuracy at 5-minute granularity, deployed on GCP with auto-retraining pipeline.

98%+accuracy
5mingranularity
12yrdata span
SARIMAXTransformersGCPDockerFirebaseReactChart.js

Face Liveness Detection System

RACKATHON 2025 Winner
Computer VisionSep 2024 — Mar 2025

Problem: Facial recognition systems are vulnerable to spoofing attacks (photos, videos, masks). Solution: Custom neural network with 477 facial landmarks + mathematical formulae for dynamic response-based liveness. Impact: 99% accuracy at <200ms latency, optimized for low-resource devices — applicable to KYC, banking, proctoring.

99%accuracy
<200mslatency
477landmarks
TensorFlowOpenCVMediaPipeFlaskDockerscikit-learn

Enterprise AI/ML Solution Suite

@scale 2026 Winner
AI / MLMar 2026

Problem: Real industry datasets with limited availability and high-performance constraints. Solution: End-to-end ML pipelines with forecasting, anomaly detection, and optimization on actual enterprise data. Impact: Won @scale AI Hackathon — evaluated by Findability Sciences / NSBT industry experts.

PythonMLEDAFeature EngineeringReactFastAPI

// all projects

LLM & RAGActive

SatSage — AI Satellite Data Query Assistant

Problem: ISRO's MOSDAC satellite data portal is difficult to query without domain expertise. Solution: Hybrid Knowledge Graph (Neo4j) + RAG architecture with FAISS vector search, fine-tuned BERT classifier (0.92 F1), and Groq LLM with streaming responses. Impact: 828 indexed documents across 12 satellites, 154 data products, geospatial mapping, and a 38-test evaluation suite.

828documents
0.92BERT F1
FastAPINeo4jFAISS+5
IoTActive

TerraByte — AI + IoT Agriculture Platform

Problem: Farmers lack data-driven tools for crop planning, irrigation, and disease detection. Solution: Unified platform with 8 sensor types, 38-class disease detection (94% accuracy), and smart irrigation. Impact: Potentially 30% yield increase and 40% water reduction.

8sensor types
38diseases
PyTorchRaspberry PiArduino+3
Computer VisionActive

SLDTS — Sign Language Translation

Problem: 1.8M+ deaf individuals in India face communication barriers. Solution: Real-time ISL sign → text/speech and speech → sign translation using MediaPipe + LSTM/Random Forest hybrid. Impact: 88% accuracy across 35 signs at <100ms latency.

88%accuracy
35signs
MediaPipeLSTMFlask+2
HealthTech

AyuUnity — Digital Health Ecosystem

Problem: Healthcare in India is fragmented — patients juggle multiple apps and portals. Solution: AI-driven unified platform with voice prescriptions, diagnostics, and ABDM/eSanjeevani integration. Impact: 6 dashboards, 4 AI models, 5+ integrations including WhatsApp Cloud API.

6dashboards
4AI models
ReactFlaskTensorFlow+3
Computer Vision

CrowdShield — AI Crowd Monitoring

Problem: Large gatherings lack real-time crowd density monitoring for safety. Solution: YOLOv8-based crowd detection with WebSocket live feeds, MQTT alerts, and geographic routing for emergency response. Impact: Real-time monitoring with automated safety alerts.

YOLOv8WebSocketMQTT+3
Computer Vision

E-Wise — E-Waste Detection + Recycling

Problem: E-waste recycling guidance is inaccessible, especially in regional languages. Solution: YOLO-based detection across 25 e-waste classes with multilingual chatbot (10+ languages) and TTS recommendations. Impact: 92% mAP, accessible recycling guidance across India.

25classes
10+languages
YOLOv8Gemini APIReact+2