15 years building systems other engineers depend on. Now designing multi-agent AI systems for production: agent state, context isolation, token cost, failure handling.
I design and build production AI systems, mostly multi-agent pipelines where the hard part is the architecture rather than the prompt: how state moves between agents, how context stays isolated so token cost doesn't grow with every step, and how a bad model output fails at the handoff instead of corrupting the whole run.
Currently building — Frezpo
Frezpo turns one client brief into a matched Google Ads + landing page strategy pack — ad architecture, landing page strategy, and design tokens generated together, so the ad's promise and the page's headline never drift apart.
The problem it targets is message match: the gap between what an ad promises and what the landing page delivers. High click-through with a high bounce rate is almost always this gap, and it opens because the ads and the page get written by different people, at different times, from different briefs.
- Product — https://frezpo.com/
- Interactive demo — https://frezpo.com/stream-demo/
| Where | What |
|---|---|
| Akbank | Refactored a hybrid React Native project into a standalone AAR package — build times 50 min → 10 min (80% reduction), removing the Java team's Node.js dependency. |
| eBay | Cut first render on a high-traffic commerce platform from 1.5 s → 600 ms by attacking virtual-DOM CPU load in SSR and moving the critical path to Node.js/EJS. |
| Appcircle | Built AI-integrated reporting over Jenkins and internal APIs, driving weekly bug tickets from 26 → 8 via feature/bug ratio analysis. |
| Microsoft & Innovance | Standardised React and React Native build/release strategy on Azure across 15 development teams, through UAT and production. |
Previously also at n11 and Takasbank.
- Personal site — https://oyilmaztekin.github.io/
- LinkedIn — https://www.linkedin.com/in/oyilmaztekin/