AI SYSTEMSlive loop
Görkem
Tanağardıgil
AI Systems Engineer
I turn messy institutional knowledge into grounded agents — support, chat, and domain specialists that answer from your data, not from thin air.
Operating model
How an agent actually runs
From knowledge base to a specialized agent — the loop people feel when support, chat, or domain bots stop guessing.
Problem · Answers float free of your docs.
Build the knowledge base
Ingest policies, tickets, manuals, and product data into a RAG store. Chunking, embeddings, and hybrid search so retrieval is specific — not a dump of the whole wiki.
Live trace
running> ingest: 1,240 docs · policies · runbooks> chunk + embed · hybrid index ready> kb: support-core · status: grounded
01
Knowledge that sticks
RAG knowledge bases so agents answer from your corpus — policies, product, ops — instead of inventing it.
02
Agents with a job
Support, chat, and domain specialists on LangGraph: clear routes, tools, and handoffs — not one mega-prompt.
03
Trust under load
Citations, evaluation, and escalation paths so production teams can ship agents without gambling on every reply.
Featured
Problems I solve
Technology is the means. These are the outcomes.
- 01
AI systems
Teams need answers from their own data — not a generic chatbot.
RAG knowledge bases → domain agents. Build retrieval-grounded knowledge bases, then specialize LangGraph agents for support, in-product chat, and domain workflows with citations and human handoff.
RAG · LangGraph · Evaluation · Tool calling
- 02
Product
Creators want faceless video without a film crew.
PostFaceless. One brief becomes narrated long-form and Shorts — reviewed by the creator, then scheduled across YouTube, TikTok, and Instagram.
TypeScript · Next.js · FFmpeg · AI pipelines
- 03
Production
Payroll cannot silently drop messages or fail under burst.
Payroll Engine. Reliable microservice payroll on Kubernetes: message-driven workflows, guarded rollouts, and systems that hold when load spikes.
Kotlin · Spring Boot · Kubernetes · MongoDB
Lab
Products & experiments
Telework Tracker
Cross-border workers lose track of telework day limits.
Track telework days on iPhone and Apple Watch against yearly caps.
Legacy Rule Extractor
Business rules are trapped in COBOL with no trustworthy extract.
AI extraction scored against a hand-written key, then replayed to confirm behaviour.
PostFaceless
Faceless channels still need production-grade pipelines.
Brief → video → creator review → schedule.
Contact
Open a channel
Roles, collaborations, or agentic product work — same channels as the contact page.