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 an AI engineer, I built an LLM evaluation platform (a 923-case benchmark, 569 cases with human-adjudicated gold labels, bootstrap confidence intervals and release gates), and I am building language models from first principles 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 · EatBetter, AI-powered nutrition app · September 2026 · case study
- Evaluation platform. Built a 923-case benchmark from scratch: 569 human-adjudicated gold labels, a 277-case sealed holdout, 16 implemented metrics and 742 automated tests in CI. Comparisons use user-clustered bootstrap CIs, with evaluators that never count a missing measurement as zero, and a three-part release gate that does not treat missing evidence as a regression.
- Model decisions by measurement. Ruled out a candidate that cost 7× more after measuring no gain (ΔF1 −0.006, 95% CI [−0.023, +0.012]), and dropped a 12× costlier candidate pipeline that trailed the bare model in 4 of 4 comparisons (exact F1, 383 DEV cases, single run).
- Assistant safety. On a frozen 400-case offline scorecard, a fix cut the missed-allergy-context veto rate from 96.4% to 2.1%, the only one of ten candidates to pass the gate. It was still behind a flag at the end of the engagement, so this is not a live incident rate.
- Production data quality. In a live PostgreSQL catalog of about 387k rows, soft-deleted 88,970 and unpublished 90,735, write steps followed plan → ROLLBACK rehearsal → expected-row-count check → human-approved COMMIT (one step skipped its separate rehearsal; the count matched).
TÜBİTAK 2209-A Sentinel-2 crop-mapping study (İzmir/Urla): built sentinel-crop-pipeline, the data-preparation stage for U-Net work (spatially blocked splits); no model results claimed.
Education. B.Sc. City and Regional Planning, İzmir Institute of Technology (expected June 2027) · B.Sc. Management Information Systems, Istanbul University AUZEF (concurrent second degree, expected June 2027).
| System | What it is | Status | Links |
|---|---|---|---|
| benchfck | Deterministic evaluation generator and exact Rust harness for machine-state tasks | v0.4.0-alpha engineering candidate · no model results published | Repo · Validity contract |
| neuron-to-gpt | Path from a single neuron to GPT-2-scale models, implemented step by step (PyTorch from week 3) | Public · Week 10 GPT-2 124M weight transfer with logits parity (historical stage) · Week 12 and G4: a separate, randomly initialised 124M-parameter model, then continued training from its own checkpoint, validation NLL 4.985 → 3.666, HellaSwag 30.5% on 1,000 held-out examples vs 33.7% for the GPT-2 reference · no claim of parity with GPT-2 | Repo |
| geoai-skills | 18 Agent Skills covering the geospatial lifecycle, routed by an orchestrator | v0.4.1 · MIT · routing benchmarked 2026-08-05 | Repo · Benchmark |
| arcgis-mcp-bridge | MCP server exposing 100 ArcGIS Pro / ArcPy geoprocessing tools to LLM hosts | Apache-2.0 · PyPI · Glama tool-definition quality A (3.5/5, scored 2026-09-25) · 8/8 smoke cases on a real ArcGIS Pro 3.7 runtime, plus 128 tests against ArcPy mocks | Repo · PyPI · Glama |
| Pixel LLM | Godot game that walks the player through building a language model in 10 stages, from data collection to launch | In development · private repo | Project page |
| PlaceTwin (product name Mendirek) | 3D digital twin and studio panel for two urban-design sites in Ayvalık, built for an IZTECH urban-design studio project, with a tool-calling Claude assistant grounded in the twin's data | Live site, member access only · private repo | Project page |
| FOUNDER.EXE | Startup simulation with Turkish and US rule sets, SAFE and cap-table mechanics | v1.7.0 Windows build ready, not yet uploaded to a store · private repo | |
| orchestration-skills | Multi-model orchestration skill set for Claude Code: 4 skills and 6 named sub-agents routing work to Claude sub-agents, the Codex CLI and Gemini | v0.1.0 · Apache-2.0 · Windows + PowerShell 7 · subscription sessions only · no measurements published | Repo |
| sentinel-crop-pipeline | Reproducible Sentinel-2 preparation: discover, mask, patch, label | v0.3.2 · Apache-2.0 · PyPI · Zenodo DOI | Repo · PyPI · DOI |
| agri-dss | Zero-backend spatial decision support across 5 districts and 147 neighborhoods in Western Antalya | Live prototype | tarimsalkoridor.online · Repo |
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.
| Arm | Precision | Recall | Full-route accuracy |
|---|---|---|---|
| Skills enabled | 99.17% | 96.77% | 96.41% |
| Skills disabled (control) | n/a | 0% | 0% |
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-gptbuilds a GPT up from one neuron and now trains a 124M-parameter model of its own; 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 (basic) · GeoPandas · Rasterio · ArcPy
Open to AI engineering work: LLM evaluation and evaluation design, agent systems, MCP infrastructure and applied machine learning. Availability and scope are best discussed directly.




