I build GeoAI agent systems, geospatial AI infrastructure, remote-sensing data pipelines, spatial decision-support systems, and applied simulation products.

My work sits between research-grade spatial analysis and usable software: standards-driven Agent Skills for reliable geospatial reasoning, ArcGIS Pro automation for LLM/MCP workflows, reproducible Sentinel-2 data pipelines, browser-native agricultural DSS tools, and realistic strategy simulations that encode real-world institutional rules into interactive systems.

AI / ML / Data Science

Spatial Data Science / Remote Sensing

Agent Systems / Evaluation

Systems / Product Engineering

An open, Agent Skills-compatible suite of 18 specialist and cross-cutting GeoAI skills that turns general-purpose AI agents into more disciplined geospatial practitioners across data acquisition, remote sensing, spatial ML, statistics, optimization, databases, visualization, and guarded ArcGIS Pro execution.

• Full-lifecycle orchestration: a geoai-orchestrator routes multi-stage work across acquisition, sensing, modeling, analysis, and delivery skills.
• Fail-loud spatial computing: explicit CRS and unit checks, anti-leakage defaults, row-count accounting, visual + numeric verification, and uncertainty as a required deliverable.
• Measured routing: a frozen 120-case, 17-skill benchmark reports 100% routing precision, 92.86% recall, and 92.5% full-route accuracy for the declared Claude Code / Claude Sonnet runtime pair; the results measure routing rather than answer quality and exclude the later ArcGIS skill.
• Evaluation infrastructure: typed scenarios, versioned JSON schemas, blind request preparation, deterministic scoring, reproducible evidence packages, and an independent-family Gemini judge adapter.
• Portable distribution: installable as a Claude Code plugin or as self-contained skills in any Agent Skills-compatible runtime.
• Guarded proprietary execution: the arcgis-pro-automation skill delegates real ArcPy work to arcgis-mcp-bridge with local runtime and mutation safeguards.

An MCP framework exposing 100 specialized ArcGIS Pro / ArcPy geoprocessing tools to LLM hosts and intelligent agents while keeping the licensed GIS runtime isolated from the host process.

• MCP-native GIS automation: ArcGIS Pro geoprocessing exposed through Model Context Protocol.
• Runtime isolation: async MCP server separated from the ArcPy worker subprocess.
• Path safety: PathGuard validation layer before filesystem and geodatabase operations.
• Distribution: published on PyPI and listed on Glama with A-rated MCP quality.
• Testing: mocked ArcPy test strategy for development without requiring ArcGIS Pro in CI.

flowchart LR
    A["LLM Host / AI Agent"] -->|MCP| B["Async MCP Server"]
    B --> C{"PathGuard"}
    C -->|validated| D["Isolated ArcPy Worker"]
    C -.->|blocked| X["Rejected"]
    D --> E["ArcGIS Pro / ArcPy"]
    E --> F["100 geoprocessing tools"]

A reproducible Sentinel-2 data preparation pipeline for crop-classification research. It prepares training-ready geospatial datasets for downstream CNN/U-Net or other remote-sensing ML models.

• CDSE-first workflow: Sentinel-2 L2A scene discovery and AOI-cropped download through Copernicus Data Space Ecosystem.
• Cloud-aware preprocessing: metadata cloud filtering, SCL invalid-pixel checks, cloud/shadow/snow masking, reflectance normalization, and spectral indices.
• Training-ready patches: COG/TIFF patches for QGIS inspection plus NPY/TFRecord exports for model training.
• Spatially blocked splits: train/validation/test assignment by spatial blocks to reduce leakage from neighbouring patches.
• Ground truth: GeoJSON/Shapefile crop polygons rasterized into patch-aligned uint8 masks.
• Research artifact: PyPI package, Zenodo DOI, CI, examples, validation run, and citation metadata.

A fully client-side Spatial Decision Support System for the Western Antalya agricultural corridor: 5 districts, 147 neighborhoods across Demre, Finike, Kaş, Kemer, and Kumluca.

• Zero-backend static architecture: Vanilla JS over a decoupled data.json layer.
• Auditable recommendation logic: compact crop and region schema resolved at runtime.
• Field-facing output: guided interface and clean A4 print plans for local agricultural planning.
• Deployment: static, easily hosted, and low-maintenance.

A browser-based startup simulation game modeling the practical stress of founding a company in Türkiye or the USA: taxes, incorporation choices, grants, investor expectations, cash flow, regulatory friction, and failure/exit outcomes.

• Rules-driven simulation: country-specific startup, tax, grant, regulatory, and runway logic.
• Idea-aware onboarding: the player writes a startup idea; the system shapes the simulation around sector and market assumptions.
• Optional AI layer: Gemini, OpenAI, or Anthropic can be connected with the player's own API key; scripted fallbacks keep the game playable offline.
• Browser-native product: static web app with local save persistence.

Downstream model-training work for agricultural pattern identification from satellite imagery. This track is separated from sentinel-crop-pipeline, which focuses only on reproducible data preparation.

A reproducible regional analysis of Türkiye's housing market before and after the COVID-19 pandemic, separating nominal changes from inflation-adjusted real price growth and using spatial statistics to detect regional clusters.

• HPI deflation workflow.
• Choropleth and comparative regional figures.
• LISA-style spatial cluster and outlier interpretation.

A reproducible urban-territorial resilience index prototype for Kaş / Bayındır, Antalya.

• Five-pillar composite indicator framework.
• Transparent normalization and weighting.
• Reproducible notebooks, figures, and methodology.

These are the narratives I am expanding into separate portfolio case studies:

• geoai-skills: turning tacit geospatial practice into portable agent instructions, then measuring whether specialist skills route correctly without overstating behavior quality.
• arcgis-mcp-bridge: separating licensed GIS execution from AI host runtimes while preserving secure, discoverable geoprocessing access.
• sentinel-crop-pipeline: making Sentinel-2 crop-classification data preparation reproducible before the model-training stage.
• agri-dss: compressing agricultural suitability and local economic knowledge into printable, village-level decision plans.
• FOUNDER.EXE: encoding company formation, tax pressure, grants, investment logic, regulatory constraints, and optional AI advising into a playable browser simulation.

• GeoAI Agent Systems: portable Agent Skills, routing evaluation, spatial-method safeguards, and reproducible evidence for agent behavior.
• Agentic GIS & MCP Tooling: secure local automation layers that expose GIS operations to LLM agents without compromising ArcGIS Pro runtime isolation.
• Remote Sensing ML Pipelines: reproducible Sentinel-2 workflows, ground-truth masks, spatially blocked splits, and downstream segmentation experiments.
• Spatial Decision Support: static, auditable, browser-native systems for agricultural and territorial planning.
• Applied Simulation Systems: strategy simulations that encode real-world regulatory, economic, and decision processes into interactive products.
• Urban & Regional Analytics: composite indices, spatial econometrics, and territorial resilience frameworks.

Open to collaboration around production-grade GeoAI agent systems, spatial machine learning, remote-sensing data pipelines, MCP infrastructure, and applied simulation products.

• Portfolio: muend.github.io
• Live App: tarimsalkoridor.online
• Game: FOUNDER.EXE on itch.io
• Email: edcoders@gmail.com