Agentic AI Engineer
Agents that plan, use tools, retrieve the right context and know when to ask a human. I connect AI workflows to real software with Python, TypeScript and C#/.NET.
Based in Istanbul, Türkiye. Working across agents, tools and applications.
Select a step to see how I think about agents.
Separate responsibilities, choose useful tools and define what a good result looks like before an agent starts working.
An illustration of my engineering approach.
Explore the systems, their source code and the thinking behind them.
Multi-agent orchestration
A LangGraph software-team prototype: architect and reviewer roles, parallel backend/frontend work, test gates, state reducers and bounded human escalation.
Implemented prototype; offline routing tests included
Reusable team configuration
Company-style roles, isolated Git worktrees and cross-model review. The public repository contains generic prompts and review policies, with application-specific context removed.
Retrieval with tool use
A personal knowledge assistant with semantic and date-based retrieval, conversation context and a model-selected MCP date resolver. The public source excludes real chat data and session credentials.
Media workflow prototype
A media-generation workflow with checkpoints, human decisions, QA and bounded provider retries. The offline path is implemented; live-provider and release validation remain separate gates.
Audio-ingestion prototype
A modular .NET audio-ingestion application with storage and noise-removal jobs, plus a mock-driven React interface. Full transcription and RAG integration remain incomplete; the screenshot shows synthetic demo data.
Model Context Protocol
My custom MCP servers connect agents to browser workflows and domain knowledge.
Job search, saved-search monitoring, events and proposal drafts. Submission stays with a human.
View source on GitHub
Browser-backed product search, filter discovery and structured results.
View source on GitHub
Offline-first knowledge search, BM25 retrieval and terminology mapping. Downloaded article corpora are not distributed.
View source on GitHub
Read-only wiki retrieval, fuzzy term matching and optional timestamped guide search. Full third-party transcripts are omitted.
View source on GitHubPersonal software foundations and a hobby project.
Personal hobby project
Warframe inventory/data tooling with a local Node.js service and a WPF desktop client. Public source excludes personal inventories, chat captures and provider binaries.
Windows desktop utility
A WPF download manager with Chrome handoff, temporary transfers and configurable category folders. Browser-only authenticated requests remain a limitation.
I work across the whole system: agent behavior, tool contracts, application code and the review loop. My .NET and automation background helps turn an AI idea into a workflow people can actually use.
Make the boundaries explicit.State, validation, retries and escalation belong in the design.
Keep a human in the loop.Approval and cancellation should be usable parts of the workflow.
Review what the agent produces.Use test gates, offline checks and cross-model review where they help.
No AI-generated posts.
A space for personal essays and technical notes I write myself.
Let's talk about agents, useful tools and the software around them.