Teamday
← People

Noyan Tokgozoglu

Head of AI Engineering at Optiver

agentsenterpriseinferenceoptiver

About Noyan Tokgozoglu

Noyan Tokgozoglu is head of AI engineering at Optiver, the global market-making firm that trades with its own capital. Before Optiver he worked at Apple. His team builds and runs Optiver's internal AI platform: model access, evaluation, agent infrastructure and the tools traders and quants use every day.

In his AI Engineer talk, "How a Trading Firm Ships New AI Models the Same Day," Tokgozoglu explains why a firm that makes more than 10 million trades a day and sees the P&L impact of a change within hours needs a modular, model-agnostic platform. His answer covers golden-set benchmarks from trading teams, an inference gateway in front of third-party, self-hosted and open-source models, a "meta harness" layer, an agent gateway, and forward-deployed AI engineers who sit with traders.

He is also candid about what isn't solved: knowledge and memory systems he is "not happy with," and evaluating agents whose tools depend on a world state, like Slack, that can't easily be replayed.

Career Highlights

  • Head of AI Engineering at Optiver
  • Previously at Apple
  • Speaker at AI Engineer (2026) on scaling AI agents and same-day model rollouts at a trading firm

Notable Positions

On Same-Day Model Adoption

Optiver aims to use a newly released frontier model in production flows the same day. That depends on a mature eval framework with benchmarks collected from key trading teams, gateway and cost-management support for the model, harness updates, round-the-clock support across regions, and clear guidance to users. See Agent Evaluation and AI Gateway.

On Staying Model- and Harness-Agnostic

All model traffic goes through Optiver's inference gateway, which handles routing, failover, budgets and logging. A meta harness layer lets teams define portable harness configurations tied to their knowledge base. He avoids vendor lock-in after an inference gateway provider put an urgent need "on the roadmap" instead of shipping it. See Agent Harness and Model Routing.

On Controlling AI Spend

The biggest leak was context: fresh sessions starting with 200k tokens or more because self-service users loaded every MCP and knowledge source. MCP tool search, CLIs wrapping MCPs, cache tuning, knowledge bases instead of huge system prompts, cheaper open-source models and self-set team limits brought spend back under control. See Context Engineering.

On Building for High-Agency Users

When traders started building their own hacky UIs, Optiver built a shared web portal for agent sessions, artifacts and collaborative documents that traders now contribute code to. He prefers scoped, versioned agents that call each other through the agent gateway over one agent that does everything.

Key Quotes

  • "If a new frontier model is made available publicly, we want to be able to leverage in the production flows the same day." (on model adoption)
  • "Sometimes people have like 200k tokens worth of context, like by default minimum, sometimes even more." (on context bloat)
  • "Our users are very high agency, they will go and build something if we don't build it for them. So we just build a better version of it." (on internal tools)