I am an AI engineer. I build LLM evaluation platforms, agent systems and the tooling around them: benchmarks with human-adjudicated gold labels and release gates, standards-driven Agent Skills, and MCP servers that give an agent real tools without handing it the runtime.

Most recently, as a contract AI engineer, I built a production LLM evaluation platform (a 923-case benchmark with human-adjudicated gold labels, bootstrap confidence intervals and release gates), and wrote a character-level GPT from a single neuron up in neuron-to-gpt. My degree is in city and regional planning, which is why much of my open-source agent tooling was first built for geospatial work.

The through-line is evidence. Every number below comes from a benchmark, a manifest or a released artifact in the repo it describes or, for client work, from the engagement's own records. Where a result does not exist yet, the repo says so instead of implying one.

AI Engineer (Contract) · EatBetter, AI-powered nutrition app · Sep 2026 · case study

• Evaluation platform. Built a 923-case benchmark from scratch: 569 human-adjudicated gold labels, a 277-case sealed holdout, 16 deterministic metrics and 742 automated tests in CI. Comparisons use user-clustered bootstrap CIs, and the release gate counts missing evidence as a failure.
• Model decisions by measurement. Retired a 7× costlier model after showing no gain (ΔF1 −0.006, 95% CI [−0.023, +0.012]) and shut down a candidate pipeline that lost to the bare model in 4 of 4 comparisons.
• Assistant safety. On a frozen 400-case scorecard, a fix for users with long memory profiles cut the missed-allergy-context veto rate from 96.4% to 2.1%, the only one of ten candidates to pass the gate.
• Production data quality. In a live PostgreSQL catalog of about 387k rows, soft-deleted 88,970 and unpublished 90,735, each write rehearsed as a ROLLBACK against a predicted row count before a human-approved COMMIT.

Undergraduate Researcher · TÜBİTAK 2209-A · 2026–2027. Sentinel-2 crop classification with U-Net segmentation and spatially blocked validation, built on sentinel-crop-pipeline.

Education. B.Sc. City and Regional Planning, İzmir Institute of Technology (expected June 2027) · B.Sc. Management Information Systems, Istanbul University (concurrent second degree).

Research repos: kutri-resilience-index, an urban resilience composite indicator with a DOI · turkiye-housing-prices-pandemic, inflation-adjusted HPI and spatial clustering · agri-unet, the downstream U-Net training track for sentinel-crop-pipeline.

geoai-skills ships a frozen routing benchmark instead of a claim about how well the skills "work". 167 cases, 18 skills, run 2026-08-05 on Claude Code 2.1.214 with claude-sonnet-5, against a paired control with the skills switched off.

The control recorded zero activations across all 167 cases, which is what makes the treatment arm mean anything. The benchmark measures routing, not answer quality, and the repo states that boundary explicitly.

arcgis-mcp-bridge has to give an agent real geoprocessing power while keeping a licensed GIS runtime out of the host process. Two isolated processes, two independent path validations, one refusal path.

flowchart TB
    H["LLM host / AI agent"] -->|MCP over stdio| S["Async MCP server"]
    S --> A["PathGuard A<br/>pre-check + confirmation gate"]
    A -->|validated request| W["Isolated ArcPy worker"]
    W --> B["PathGuard B<br/>independent re-validation"]
    B -->|validated path| G["ArcGIS Pro / ArcPy runtime<br/>100 geoprocessing tools"]
    G --> O["Structured NDJSON result"]
    A -.->|blocked| X["Rejected safely"]
    B -.->|blocked| X

    style H fill:#ECF2EC,stroke:#1A221C,color:#1A221C
    style S fill:#D7E9DD,stroke:#0B7F47,color:#1A221C
    style A fill:#F6E6C8,stroke:#8A5A11,color:#1A221C
    style W fill:#DCE5EE,stroke:#2C67A0,color:#1A221C
    style B fill:#F6E6C8,stroke:#8A5A11,color:#1A221C
    style G fill:#D7E9DD,stroke:#0B7F47,color:#1A221C
    style O fill:#ECF2EC,stroke:#1A221C,color:#1A221C
    style X fill:#F3D9D4,stroke:#A83A2B,color:#7A2418

• Deterministic evaluation. Exact generators, restricted verifiers, preregistered release gates, and a hard line between engineering evidence and model evidence.
• Language models from first principles. neuron-to-gpt builds a GPT up from one neuron with a measured six-stage comparison; Pixel LLM turns the same pipeline into a game.
• Agent systems. Turning tacit expert practice into portable agent instructions and guarded tools, then measuring whether the routing actually holds up.
• Remote sensing ML. Spatially blocked splits, patch-aligned ground truth, and segmentation experiments downstream of a pipeline that is reproducible first.

Stack: Python · TypeScript · Rust · PyTorch · FastAPI · PostgreSQL / PostGIS · MCP · Agent Skills · Docker · GeoPandas · Rasterio · ArcPy · PySAL

Open to AI engineering work: LLM evaluation and evaluation design, agent systems, MCP infrastructure and applied machine learning. Available part-time now and full-time from July 2027.

Portfolio (Türkçe) · LinkedIn · nsduraan@gmail.com