Software developer and Management Information Systems student in Istanbul. I build Python tools that make AI workflows easier to test: paired evaluation comparisons, agent trace linting, RAG citation checks, MCP contract tests, and reproducible API replay.

In team projects, I work across FastAPI backends, human-in-the-loop decision systems, Gemini integrations, and read-only data tooling.

Portfolio / LinkedIn / Email / All repositories

_{Generated weekly from the GitHub API. Forked and archived repositories are excluded from the language mix.}

Developer in YZTA Bootcamp 2026 AI Track Group 60. I contributed five merged pull requests spanning the cost summary API, interactive dashboard and themes, CI fixes, bilingual documentation, market analysis, and UI consistency.

Cost API, PR #1 / Dashboard, PR #4 / CI and docs, PR #5 / Market watch, PR #6 / UI consistency, PR #8

_{CloudSentinel night-mode dashboard, introduced in merged PR #4.}

YZTA AI Hackathon Team 215 project. Across 25 commits, I built the Gemini backend layer, read-only SQLite MCP tools, model failover and quota backoff, diagnostics, tool orchestration, and query guardrail tests. The merged backend PR changed 13 files with 615 additions.

Inspect merged PR #1

eval-delta compares paired evaluation runs by case and metadata slice, then produces seeded bootstrap confidence intervals plus terminal, JSON, and Markdown reports with CI-ready exit codes.

agent-trace-lint checks agent tool-call traces for protocol errors, invalid arguments, exposed secrets, risky shell operations, loops, and latency. Reports are available as text, JSON, or SARIF.

rag-citecheck checks RAG evaluation records for missing citations, unknown sources, token overlap, and absent quoted evidence. It also emits terminal, JSON, or SARIF output.

mcp-probe runs stdio MCP contract scenarios across schema discovery, tool calls, assertions, and latency, with JSON and JUnit reports for CI.

llm-replay-proxy is a FastAPI proxy for recording and replaying non-streaming OpenAI-compatible POST calls with local JSON cassettes.

Chocolate Ratings Dashboard explores 2,530 chocolate reviews in an R flexdashboard using ggplot2, Plotly, and DT, published with GitHub Pages.

For project, internship, or collaboration conversations, email mertk.mek@gmail.com or visit mertefekurt.me.