# Salih Küçükkayış

> Full-stack and forward deployed engineer in Istanbul. I take a problem from the first conversation to production — lately, mostly with LLMs.

Full-Stack Developer · Forward Deployed Software Engineer. Istanbul, Türkiye. Contact: skucukkayis@gmail.com.

## Now
- Building my own product: Feedwatch, under Saliga Software — written, shipped and supported end to end by me.
- Shipping LLM-backed features: API integrations, retrieval, structured output, agent tooling.
- Picking up Go: Mostly for small services and CLI tools.
- Open-sourcing the tooling: Small, tested tools around MCP, evals, structured extraction and retrieval.

## Work
- [BilginPro](https://bilgin.pro) (Now): A software company building high-traffic news platforms, web and mobile apps for Turkish media since 2004.
- [Saliga Software](https://saligasoftware.com) (My studio): Where my own products live. Feedwatch is the first one.

## Projects
- [Feedwatch](https://feedwatch.co): Finds the products Google and Meta silently refuse to list, prices the gap at the merchant's own list prices, and fixes it from the same screen. Stack: TypeScript, React Router, PostgreSQL, Prisma, GraphQL, Docker.
- [catalog-mcp](https://github.com/msalihk/catalog-mcp): MCP server that audits a product catalog for Google Shopping and Meta readiness: per-rule breakdown, sample products, affected catalog value. Stack: TypeScript, MCP, Zod.
- [laravel-mcp](https://github.com/msalihk/laravel-mcp): Turn a Laravel app into an MCP server. Exposes only the tools and Eloquent columns you list, read-only, with Gate checks. Stack: PHP, Laravel, MCP.
- [pg-readonly-mcp](https://github.com/msalihk/pg-readonly-mcp): MCP server in Rust that lets an assistant explore PostgreSQL without being able to change it — SQL parsed and checked before it runs. Stack: Rust, PostgreSQL, MCP.
- [agent-evals](https://github.com/msalihk/agent-evals): Small eval harness for LLM features and tool-calling agents. Cases in TypeScript, run in CI, the build fails on a regression. Stack: TypeScript, OpenAI, Anthropic.
- [structured-extract](https://github.com/msalihk/structured-extract): Messy product titles into validated, typed feed attributes with an LLM, then deterministic checks that discard what can't be verified. Stack: Python, Pydantic, OpenAI, Anthropic.
- [newsroom-rag](https://github.com/msalihk/newsroom-rag): Question answering over a Turkish news archive with citations, hybrid search, and a refusal when the archive doesn't support an answer. Stack: Python, FastAPI, PostgreSQL, pgvector.

## Stack
- Languages: TypeScript, Python, Rust, PHP, Go
- Backend: Laravel, Node.js, FastAPI, Redis
- Data: PostgreSQL, MySQL
- Frontend: React, Vue, Nuxt, Inertia, Tailwind CSS
- AI / LLM: OpenAI API, Anthropic API, MCP, Vector search
- Infra: Docker, Linux, GitHub Actions

## Writing
- Nothing published yet.

## Links
- GitHub: https://github.com/msalihk
- LinkedIn: https://www.linkedin.com/in/salihkucukkayis
- X: https://x.com/skucukkayis