Projects
Research and engineering projects across AI, CV, NLP, and systems.
NLP & LLM
LILAI — Grammar & Phonology Induction with LLMs
Investigated grammar and phonology induction in low-resource languages using in-context learning with Qwen2.5-14B-Instruct. Designed inference strategies focused on calibrated prompting and minimal architectural complexity, achieving chrF 0.2998 on the IOL-AI 2026 benchmark. Demonstrated that careful prompt engineering with a single forward pass can match or exceed complex multi-stage approaches.
ChatGS — Multi-Modal Geospatial AI Assistant for NRSC/ISRO
Unified text, voice, and sign language interface for satellite data querying, reservoir monitoring, and ISRO resource lookup. Integrates Google Earth Engine with Gemini, Groq, and Mistral LLMs for real-time geospatial analysis. TensorFlow-based ASL recognition enables accessible interaction. Deployed as a Flask web application at NRSC Hyderabad, supporting reservoir health assessment, satellite footprint queries, and multilingual voice interaction.
AutoGEO — Conversational Oceanographic Data Analysis Platform
Architected and guided student development of a natural language query engine over Google Earth Engine for oceanographic research covering sea surface temperature, ocean currents, and sea level analysis via plain English prompts. Platform powered by Groq/Llama 3.3 70B and Mistral AI with PostgreSQL-backed query history and publication-ready visualizations through Geemap and Folium.
Chat with PDFs — RAG System
Built a multi-document question answering pipeline combining LangChain for retrieval orchestration, OpenAI embeddings for semantic search, and FAISS for efficient vector indexing over document chunks. Deployed on Hugging Face Spaces for interactive document querying and natural language answers.
Computer Vision & Remote Sensing
River Morphology & Water Quality Analysis at Triveni Sangam for Mahakumbh-2025
Multi-sensor remote sensing study of the Ganga river at Prayagraj before, during, and after Mahakumbh-2025 using Sentinel-1 SAR (morphology) and Sentinel-2 MSI (water quality). Quantified 2 m channel deepening from dredging, stream unification into 250 m wide channel, and dissolved oxygen fluctuations from 10.2 to 6.3 mg/L during mass bathing events. Findings informed UP Irrigation Department flood management.
SkyServe — Vehicle & Ship Detection (Production)
Production-grade computer vision models for remote sensing, including high-precision convoy and ship detection. Developed a novel Shape-Aware OBB to HBB conversion algorithm improving mean IoU from 0.4929 to 0.5609 on ShipRSImageNet. Includes real-time convoy detection and automated road health assessment pipelines.
LWH — Large-Scale Ship Detection & Heading Estimation
Multi-sensor satellite imagery pipeline for ship detection, heading estimation, and dimensional measurement across Sentinel-1 SAR, Sentinel-2 optical, and PlanetScope sensors. Hybrid heading estimation combining border wake analysis, NIR wake detection, and ML-based prediction. Trained sensor-specific detection models with confidence scoring, deployed operationally at SkyServe for maritime surveillance.
Ghost Ship Detection — SAR-based Maritime Anomaly Detection
Production pipeline detecting vessels with disabled AIS transponders by correlating YOLOv11 detections on SAR satellite imagery against live AIS signals. Features geospatial visualization dashboard built with Streamlit, multi-sensor support (Sentinel-1, RADARSAT), and automated ghost ship flagging. Developed during ISRO/NRSC internship for maritime security.
Stubble Burning Prediction with xLSTM from Multi-Sensor Satellite Data
Deep learning framework integrating NASA VIIRS, MODIS, and Sentinel-1/2 data for stubble burning detection and forecasting. xLSTM neural network achieves predictive modeling of fire events across agricultural regions. Includes comprehensive EDA, temporal correlation analysis, and interactive HTML map outputs for operational monitoring.
Solar Panel Detection & Power Estimation
Detected solar panel installations from Sentinel-2 and WorldView-3 satellite imagery using a YOLOv11 object detection model. Estimated power generation potential by combining detected panel surface area with NASA POWER solar irradiance data for the corresponding geographic region.
Road Network Detection from Satellite Imagery
Conducted a comparative evaluation of road extraction methods from satellite imagery, testing classical edge detection, morphological operators, and machine learning approaches including Random Forest, SVM, and deep learning. Benchmarked on the DeepGlobe and SpaceNet datasets to identify the most effective approach for automated road network mapping.
Lip Reading with LipNet
Reproduced the LipNet architecture using 3D convolutional layers for spatiotemporal visual feature extraction combined with bidirectional LSTMs for temporal sequence modeling. Trained on the GRID corpus using Connectionist Temporal Classification loss for sentence-level lip reading from video.
GAN Artistic Style Transfer
Implemented a generative adversarial network for artistic style transfer using VGG19-based perceptual loss to preserve content structure and Gram matrix matching to encode style textures. The generator learns to transfer artistic characteristics from reference paintings to input photographs.
Real-Time Exercise Form Tracking
Built a real-time exercise form analysis system using MediaPipe pose estimation to extract 33 body landmarks from webcam video. Applied cosine similarity scoring between user pose and reference pose sequences to detect form deviations during exercise execution.
Traffic Road Sign Detection — YOLOv5
Fine-tuned a YOLOv5 object detection model on a custom dataset of four traffic road sign classes captured under varying lighting and weather conditions. Achieved real-time detection performance suitable for deployment in driver assistance systems.
Machine Learning
Atrial Fibrillation Detection from ECG
Compared neural network architectures against classical scikit-learn classifiers for atrial fibrillation detection from ECG signals. Applied SMOTE oversampling to address class imbalance in arrhythmia datasets and evaluated precision-recall tradeoffs across models.
DeepVacPred — Vaccine Subunit Prediction
Designed a four-layer deep neural network for in-silico screening of vaccine subunit candidates. Used Z-descriptor encoding to represent amino acid sequences and trained the model to predict immunogenic potential, accelerating candidate selection before wet-lab validation.
Stock Price Prediction — LSTM vs Prophet
Built a comparative stock price forecasting platform allowing users to select any ticker symbol and compare LSTM-based deep learning predictions against Facebook Prophet time series forecasts. Deployed as an interactive Streamlit application with visual performance metrics including MAE and RMSE.
Systems & Infrastructure
chemInteractions — Chemical Reaction Prediction Engine
Dual-brain reaction prediction engine fusing a 1.2B-parameter ReactionT5 Transformer (ONNX-optimized in Rust) with textbook functional group matching from a curated knowledge base. Delivers explainable predictions with confidence tiers from pure ML inference to rule-verified results. Deployed as a high-performance axum REST API with SQLite-cached PubChem enrichment.
ISRO Autonomous Robot — Path Planning, Web Control & Onboard Firmware
Full-stack robotics platform encompassing A* and RRT* path planning simulation, a Django-based web control dashboard with keyboard/audio input, and Raspberry Pi onboard firmware integrating motor control (PCA9685), QR scanning, voice commands, face detection, and IR obstacle avoidance. Built for NRSC/ISRO physical bot deployment.
SSH Portfolio — Interactive CLI on badha.is-a.dev
Interactive SSH-based portfolio server written in Rust, presenting a full research showcase through a terminal interface. Implements async I/O, raw terminal rendering, and keyboard navigation accessible via ssh ssh.badha.is-a.dev -p 2222.
Homelab Infrastructure Automation
Designed and deployed a reproducible, containerized research environment using Proxmox for VM orchestration, Docker for application containerization, and Tailscale for secure mesh networking across devices. Provides isolated development and testing environments for reproducible research workflows.
2D-to-3D NFT Model Generation
Built a pipeline converting 2D artwork into 3D mesh models using monocular depth estimation to generate depth maps followed by surface reconstruction algorithms. Produced renderable 3D assets from flat digital illustrations for NFT and 3D printing applications.
ESP32 Wi-Fi Beacon Spam
Implemented raw 802.11 frame injection on ESP32 microcontrollers with channel hopping to broadcast probe request frames across multiple Wi-Fi channels. Explored low-level network protocol manipulation for wireless experimentation and security research.
Ethereum Smart Contract — SimpleStorage
Developed and deployed a SimpleStorage Solidity smart contract to the Ethereum Sepolia testnet using Web3.py for transaction management and Brownie for contract compilation, testing, and deployment automation.
Mobile & Full-Stack
BusYatra — VIT Bus Tracking System
Developed a multi-application bus tracking ecosystem for VIT campus with separate rider, driver, and QR scanner applications. Riders view real-time bus locations, drivers receive optimized route information, and QR-based boarding verification ensures accurate ridership tracking.
Photon — Lemmy/Fediverse Client
Built a progressive web application frontend for the Lemmy Fediverse platform using SvelteKit with server-side rendering for fast initial page loads. Implemented internationalization support for multiple languages and offline-capable PWA functionality for mobile access.
Eme Safety App
Built a women's safety mobile application with one-tap emergency alerts, automatic location sharing with preselected emergency contacts, and direct integration with local helpline numbers. Designed for quick activation with minimal friction during emergency situations.