DhatuVerse

Overview

DhatuVerse is a mineral exploration platform built at NCAIR, IIT Bombay, where I work as a Project Research Assistant after joining as an ML Research Intern. The platform ingests the raw material of exploration — borehole Collar/Survey/Assay tables, LiDAR scans, terrain models and multispectral satellite imagery — and turns it into ore estimates, mineralised-zone delineations and interactive 3D subsurface views that geologists can interrogate directly in the browser.

I designed the UI/UX in Figma from scratch and built the full stack behind it: the client-facing React frontend, a Node.js/Express API server, and a Python FastAPI microservice for ML and geological computation.

Key Features

• Stereological ore estimation and multimodal AI-driven EDA that automates pattern discovery, anomaly flagging, and volume/grade estimation across heterogeneous geological, assay and survey datasets — replacing a manual interpretation workflow that previously took geologists 2–3 hours per site survey.
• Borehole desurvey and compositing pipeline built from scratch, converting raw Collar/Survey/Assay CSVs into 3D spatial trajectories across 39 collars, with length-weighted assay compositing and lode calculations to delineate mineralised zones.
• In-browser subsurface visualisation on Potree and 3D Tiles, rendering large-scale LiDAR and terrain scans for real-time interactive geological interpretation.
• Fault-tolerant chunked ingestion streaming 100+ GB of LiDAR, tabular and geospatial data with built-in error recovery, schema validation and resumable chunk retries across multi-gigabyte production loads.
• Copernicus multispectral integration wired into the core ingestion pipeline to power bauxite deposit prediction over user-selected bounding boxes, extending the platform beyond borehole-only datasets.
• Production infrastructure and CI/CD on AWS EC2 with Nginx, Docker, Postgres and S3, plus a GitHub Actions pipeline that promotes releases across staging and production using blue-green deployment for zero-downtime cutovers and instant rollbacks.

Impact

Interpretation work that used to consume hours per site now runs as an automated pass over the full dataset, and the results are explorable in 3D rather than locked in spreadsheets. On the infrastructure side, hardening four core subsystems — frontend, API layer, database and object storage — with container health checks, connection pooling and automated nightly backups cut idle compute cost by 70% while making deployments routine.