> projects

Things I've built.

Production systems, research projects, and things that started as experiments and became real.

> featured

vibeapp

in developmentsolo project

AI-powered iOS app that lets users describe any web app in plain language — or voice — and instantly deploys it live on Vercel. Streaming Claude chat with vision, Groq Whisper transcription, and one isolated Vercel project per generated app. Built solo — backend, mobile, infra, everything.

stack
Next.js 15React NativeExpo SDK 55Drizzle ORMNextAuth v5Claude Haiku 4.5Groq WhisperVercel Deploy APISSE Streaming
key decisions
Claude Haiku 4.5 over GPT
Streams faster at lower cost for code generation. Latency matters when the user is watching an app get built in real time.
Next.js + Node SDKs over Python backend
Anthropic, Groq, and Vercel all have first-class Node SDKs. One runtime, one deploy target, no cross-language glue.
One Vercel project per generated app
Each user-generated app gets its own isolated Vercel project via the Deploy API. Clean separation, independent logs, no shared blast radius.
GitHub — private during developmentiOS app — in development
product

StackFit

in developmentsolo project

A tech-stack matchmaker for developers. Answer a short quiz about your goals and constraints, and StackFit recommends a curated stack and learning path for your archetype — from “vibe coder” to full-stack engineer. An interactive catalog maps tools, ecosystems and roles, on a dark terminal-inspired design system.

stack
Next.js 16React 19TypeScriptTailwind 4shadcn/uiZodVitesti18n (TR/EN)
> production work

RAG Pipeline — Design Dataset Search

floorplannerjunior ai developer · 2025

Built a production RAG pipeline over Floorplanner's large-scale design asset dataset — millions of rooms, floors, and product placements. Designed the retrieval architecture end-to-end so the LLM could actually reason over the data instead of just searching it.

what I built
The Chunker
Internal tool that transforms Floorplanner's design database into structured chunks optimized for LLM consumption and vector search. Became the canonical way to feed design data to LLMs.
Hierarchical chunking strategy
Project → floor → room → items. Preserves structural relationships so retrieval returns contextually coherent results instead of orphaned fragments. Major retrieval-quality improvement over flat chunking.
Semantic search + product-in-room prototypes
Prototyped ML models for semantic search across design metadata and AI-assisted product placement inside rooms. Directly improved usability of the design dataset for downstream AI features.
stack
PythonLangChainQdrantEmbeddingsHierarchical ChunkingSemantic SearchVector SearchRAGLLMs
open source

loop-engineering

claude code skillgithub.com/selmakcby/loop-engineering

A drop-in Claude Code skill for building a self-running agent loop — one that keeps working until a goal is met instead of the manual write → wait → review → re-prompt cycle. The skill runs an honesty test, picks manual vs automatic and open vs closed, and wires up a closed loop with a real, un-foolable verification gate (a Stop hook running actual tests — zero tokens) plus a turn/budget ceiling. The one rule: never let the AI verify its own "done."

stack
Claude CodeLoop Engineering/goalStop HookClosed LoopModel RoutingMIT
> other work