muend.
How I build

Practice

The same sequence governs infrastructure, research software, decision systems and simulations. It is ordered so that the expensive mistakes are made cheap.

StagesFour, always in order
Primary domainAI engineering and evaluation
Failure modeSilent, unmeasured error
Phase 01

Define invariants

Make CRS, units, data contracts, safety boundaries and success criteria explicit before implementation.

Phase 02

Build the pipeline

Separate components, isolate risky runtimes, preserve provenance, and make intermediate states inspectable.

Phase 03

Stress the failure modes

Test negative cases, leakage, invalid paths, ambiguous routing, missing data and operational boundaries.

Phase 04

Ship a usable artifact

Deliver a package, live interface, citable release, benchmark, print output or deployment-ready system.

Principles

What that sequence is protecting against

01 / Measurement rigor

Assumptions are explicit

Units, coordinate systems, label provenance and the boundaries of each sample are checked before a method is trusted.

02 / Honest validation

Leakage is treated as a risk

Dependent data requires blocked or clustered splits, sealed holdouts, appropriate baselines and validation designs that reflect deployment.

03 / Reproducibility

Evidence is versioned

Typed contracts, CI, immutable benchmark packages, test harnesses and DOI-backed releases keep work reviewable.

04 / Product utility

Methods become usable systems

Analysis is translated into interfaces, reports, local automation or playable simulations rather than left as notebooks.

Stack

Technical core

A spatial-first stack spanning research methods, agent infrastructure, production engineering and browser-native product work.

Agents / evaluation

Tool-using agents that route, execute and fail safely.

Agent Skills, MCP, runtime isolation, typed evaluation contracts, blind request preparation, deterministic scoring and evidence-aware benchmark publication.

Primary domainAgent systems and LLM evaluation
Core failure modeUnsafe or wrong routing
DeliveryPlugin · MCP · benchmark
Working set10
Agent SkillsMCPJSON-RPCJSON SchemaRouting evalsBehavior evalsGuarded executionBenchmark harnessesLLM-as-judge calibrationBootstrap CIs
Fig. 01

How the projects feed each other

The work compounds: methods inform infrastructure, infrastructure enables execution, and both feed products that can be used outside a research notebook.

Fig. 01 / System atlas
SYSTEM ATLAS / KNOWLEDGE → EXECUTION → ARTIFACT REV. 04 / 2026 ROUTING / INVARIANTS GUARDED EXECUTION DATA / MODEL EVIDENCE PUBLIC DELIVERY CORE PRACTICE Reliable spatial systems RESEARCH RIGOR · ENGINEERING · DELIVERY N-01 / EVALUATION + KNOWLEDGE benchfck + geoai-skills Exact benchmarks and routed invariants; EatBetter applies them in production. N-02 / EXECUTION LAYER arcgis-mcp-bridge Runs licensed ArcGIS operations behind guarded process boundaries. N-03 / DATA + MODEL LAYER Sentinel pipeline + U-Net Reproducible EO datasets, segmentation, and neuron-to-gpt for model internals. N-04 / PRODUCT LAYER agri-dss · FOUNDER.EXE With PlaceTwin and Pixel LLM: knowledge turned into usable interactive systems.
Open to collaboration

Build AI systems that hold up.

© 2026 Muhammed Enes Duran Static site · no trackers
Technical diagram