My Projects

Comprehensive research into AI-powered Earth observation and satellite imagery analysis, focusing on TerraMind GFM and AlphaEarth technologies. This master's thesis explores cutting-edge developments in geospatial foundation models, multimodal data processing, and advanced machine learning architectures for satellite imagery interpretation.

Self-hosted web platform for satellite image interpretation and land-cover validation, built with Xiao Tan as a lighter, faster alternative to Collect Earth Online. Multi-source high-resolution basemaps, 10 built-in LULC schemas and three assessment modes (point, pixel/plot, grid/cell), with every project stored as one portable JSON file on disk. It produced the independent human reference labels used for the MSc thesis accuracy assessment.

A comprehensive professional Python library for PlanetScope satellite imagery discovery, analysis, and processing. Designed for researchers, GIS analysts, and Earth observation professionals seeking advanced satellite data workflows.

Pan-India cadastral web platform that unifies land records scattered across 28+ separate state portals into a single mobile-first map: click any plot and see owner, ownership history, area, land type and source portal, in English, Hindi or Urdu. Built as a sprint-driven solo product over three days — a pluggable state-portal scraper framework, a Prisma/PostGIS schema with row-level security and audit logging, MapLibre mapping and GeoJSON/KML/PDF export. Deployed at land.trenlens.com; paused after sprint 7 on a government data-access blocker.

Developed Python-based risk analysis software for landslide monitoring. Utilized AWS relational database based on PostgreSQL with PostGIS extension for spatial data management and analysis.

A Python library for geospatial raster enhancement through histogram manipulation. Offers stretching, matching, and equalization modules to improve image contrast and normalize datasets. Built with Rasterio and NumPy, it preserves spatial integrity while providing visualization tools for comparing original and processed imagery, essential for remote sensing applications.

In our project, we conducted a comprehensive analysis focusing on landslide susceptibility in Northern Italy. Our mission was to enhance terrain stability awareness and improve risk management through detailed mapping.

A comprehensive web application that provides support, resources, and education to victims of domestic violence, their families, and the public. The platform serves as a safe place where individuals can access critical information, connect with support services, and engage in community discussion.

Cross-platform, local-first desktop AI assistant that orchestrates arbitrary tools over the Model Context Protocol. A Rust/Tauri v2 host owns the local SQLite database and all secrets while a Next.js static-export frontend reaches them only through a typed IPC bridge; BYOK provider keys are sealed with AES-256-GCM under an OS-keychain master key and never cross into the webview. Its standout subsystem is an end-to-end-encrypted "authenticated blind relay" that lets an installable mobile PWA drive the desktop engine. Publicly released under Apache-2.0 at 0.1.0-alpha.4.

End-to-end automated systematic-review system implementing the PRISMA 2020 workflow: multi-database search, DOI + fuzzy-title deduplication, a rule-based keyword screening pass and an AI-assisted eligibility pass driven through MCP, ending in an interactive PRISMA 2020 flow diagram. One JSON state model is exposed through four front doors — CLI, FastAPI service, Next.js dashboard and an MCP server. Built to run the MSc thesis literature review, then generalised and released under MIT.

Visual pipeline builder for data science: place and wire module blocks on an infinite canvas, and an agentic layer generates, validates and runs the corresponding Python inside a Docker sandbox — including building a whole pipeline from a plain-English prompt. Implemented as a Turborepo/pnpm monorepo (React 18 + Vite + React Flow front end, Fastify 5 + Prisma 7 API, seven shared packages) and developed in explicit phases 0 through 2B with ADRs and written implementation guides. Phase 2B is merged and released as v0.6.0, with the canvas to Python to sandbox chain verified end to end.

Computer-vision pipeline that analyses bowling videos end to end. Phase 1 detects all four lane boundaries (HSV lane masking, connected-component gap filling, context-aware scoreboard/logo removal, Sobel gradients and angle-constrained MSAC/RANSAC line fitting); phase 2 tracks the ball through motion detection, blob analysis, homography rectification, RANSAC trajectory fitting and rolling-median post-processing with overlay video output; phase 4 counts toppled pins by frame differencing and contour analysis. Built as a pair project alongside the Image Analysis and Computer Vision course, which was taken as an extra course.

Full-stack artisan marketplace (Nuxt 4 / Vue 3 / MongoDB) connecting Indian artisans with customers, framed around UN SDG 8. Built end to end for a family member: dual seller/customer accounts, product catalogue and orders, and a seller analytics dashboard.

Android app (Flutter + Firebase) that helps Indian citizens discover, match against and apply for central and state government welfare schemes. Ships a 41+ scheme database across 9 categories with a 0-100% eligibility-scoring engine, a personalised feed, and the full application lifecycle from document upload to status tracking. Built end to end for a family member.

Cross-platform desktop application for AI-powered audio and video transcription, now at v2.0.0 "Turbo Edition". v2 replaced openai-whisper with faster-whisper/CTranslate2 and added batched GPU inference, INT8 quantisation, Silero VAD silence-skipping and greedy decoding, for a measured ~50x real-time speedup. A split-pane Tkinter monitor streams segments as they are produced with confidence colouring, live word count and ETA; outputs cover plain text, detailed transcripts and SRT/VTT subtitles across 50+ languages. MIT-licensed and public on GitHub.

Self-hosted platform that turns a single Linux server into a shared multi-user cloud, handing out CPU, RAM, GPU and disk to several users through JupyterHub notebooks and Docker/Portainer containers. Its core engineering idea is a Distribution Abstraction Layer: six shell libraries that let the rest of the codebase call the same install/service/path/resource APIs across six Linux distributions, systemd and SysV init, and cgroups v1 and v2. Phases 0-2 of a six-phase plan are complete.

An AI-native portfolio platform built with Nuxt, Vue and Supabase. Beyond the public site and admin dashboard, it runs two production agents on one provider-agnostic core: a private assistant that manages content conversationally with approval-gated edits, and a public, topic-locked visitor assistant that answers questions about my work from published career data and reads my public GitHub live. A private career database is the single source of truth, so the projects, the CV page and the assistants never disagree.

Fresh Flow is a Flutter-based mobile application leveraging Firebase for real-time database management and Google ML Kit for barcode and text recognition. It enables efficient grocery inventory tracking, expiry date monitoring, and dynamic shopping list generation. Core functionalities include authentication, CRUD operations, cloud synchronization, push notifications, and modular UI customization.