Geospatial Data Scientist working across LiDAR, aerial imagery, computer vision, and agentic AI systems — from field survey through production platforms supporting critical utility infrastructure.
I work at the intersection of geospatial engineering, applied AI, and the software/platform work that ties them together — currently as a Data Scientist at The Spear Group, supporting Florida Power & Light's FPLAir/NextVision program, where I lead AI/ML image recognition workflows for power equipment assessment and build the geospatial data pipelines, infrastructure-as-code, and internal platform tooling behind them.
My background started in the field: control and topographic surveys, GNSS and UAV data collection, and geodatabase design in Nepal, before an MSc in Geosciences (GIS concentration) at Florida Atlantic University sharpened the AI/ML and geospatial data engineering side. That combination — knowing how spatial data is actually collected and knowing how to build production systems on top of it — is the throughline across everything I build.
Day to day that means moving fluidly between layers: writing the Python/PyTorch models, designing the AWS CDK infrastructure they run on, building the FastAPI/React consoles that operate them, and — increasingly — building LLM agents that let people query all of it in plain English. I'm FAA Part 107 certified and hold graduate research contributions in flood hazard exposure modeling.
Two self-directed builds applying LLM agents and RAG to real problems — documented honestly, including the bugs I hit and how I fixed them. Both run entirely locally via Ollama: no API keys, no usage cost.
A natural-language interface over spatial/infrastructure data. Ask a question in plain English about San Francisco's utility infrastructure — a local LLM agent (LangGraph ReAct, Ollama llama3.1:8b) translates it into a spatial SQL query, runs it, and answers with a natural-language summary and an interactive map. Seeded with ~1,500 real utility assets pulled live from OpenStreetMap's Overpass API.
A LangChain-based RAG chatbot that indexes local documents into a Chroma vector store and answers questions in a context-aware, multi-turn conversation — grounding every answer in retrieved source material and citing which files it used. Runs entirely locally via Ollama.
Production AI needs infrastructure, data pipelines, and internal tooling around it. This is a selection of the platform-layer components I've built to support enterprise geospatial and AI systems.
AWS CDK component framework with enterprise naming, tagging, and deployment abstractions for ML and app workloads.
Modular platform for taxonomy alignment, schema publishing, quality validation, and operational patch workflows (Python, SQL, DuckDB).
CLI and web console for environment-aware API operations with shared authentication and endpoint tooling (FastAPI, React, Typer).
PostgreSQL schema migration CLI with source/target connection separation, dry runs, and preflight checks.
ONNX-based object detection package for utility equipment with COCO JSON output and optional visualization.
Reusable package for API extraction, taxonomy joins, metadata patching, and downstream publishing automation.
Boundary-focused architecture separating API adapters, data sync logic, and publishing interfaces from generic scripts.
Modular script layer for analysis, checks, loaders, fixes, and patch orchestration with migration-safe wrappers.
Optional FastAPI sidecar for heavier geospatial processing, conversion runtimes, and algorithm endpoints.
Focused API exposing health, pagination, and entity summary endpoints for analytics demos and stakeholder review.
Local map and API demo environment for assignment workflows, status transitions, and pre-publish validation.
Jupyter integration embedding the full GIS app with two-way project state sync for notebook-driven workflows.
React/TypeScript (Vite) frontend for validating operator lifecycle actions before production integration.
Storm and pilot deployment dashboard monorepo combining ArcGIS runbooks, SQL and shell automation, and demo APIs.
Data science workspace scaffold for experimentation, feature engineering, and model lifecycle workflows.
Structured runbooks, standards, acceptance gates, and deployment checklists for end-to-end delivery governance.
A small browser-side terrain model in the spirit of the FEMA-based hydrologic exposure work below — procedurally generated elevation, not the published dataset. Drag the slider to raise sea level and watch inundation spread.
Color ramp: low-lying green → upland tan/white. Blue = below current sea level.
Utility corridor: 0 of 7 poles below sea level
Most of my professional work is confidential to the organizations I've supported. Below is research and academic/course work I can share in full.
Land cover classification and change prediction using Google Earth Engine.
View on GitHub ↗Intensity image generation and DSM/DEM derivation from LAS datasets.
View on GitHub ↗Flight planning, image capture, and processing for a full campus orthophoto — Pashchimanchal Campus, Nepal.
View on GitHub ↗Footprint mapping in flood-prone regions using OpenStreetMap and HEC-RAS hydrologic modeling.
View on GitHub ↗Census-tract-level 2016 turnout prediction for Florida using forest-based classification and regression.
View on GitHub ↗Suitability modeling for low-carbon-footprint corporate site selection, Hardee County, FL.
View on GitHub ↗MSc, Geosciences — GIS Concentration · Florida Atlantic University · 2023–2024 · GPA 4.0/4.0
BSc, Geomatics Engineering · Tribhuvan University, Nepal · 2016–2021 · GPA 3.64/4.0
Open to Data Science, GIS/Geospatial Engineering, and Data/Platform roles. Based in West Palm Beach, FL.