How Optiver ships new frontier AI models the same day, with spend control
AI Engineer · youtube · 2026-10-27
In an AI Engineer talk, Noyan Tokgozoglu, head of AI Engineering at market maker Optiver, explains how his team adopts AI at trading speed.
Key points:
- Hours, not quarters: a change rolled out to traders shows up in P&L within hours, so AI iteration must move just as fast and stay measurable.
- Same-day model adoption: new frontier models are benchmarked on internal golden sets before rollout, enabling day-one shipping.
- Platform stack: an on-prem AI platform built around an inference gateway handling routing and budgets.
- Model- and harness-agnostic "meta harness": a unified layer decoupled from specific models and eval harnesses, so switching models requires no rewiring.
- Spend control: after this summer's wave of big model releases pushed AI costs up, gateway-level budgeting and usage controls brought spend back under control.
- Adoption: embed AI where traders already work rather than building separate entry points.
The talk runs 17 minutes and ends with Q&A.
More from coding & agent
- Running AI agents on a legacy IBM WMS: enforce tool order, script checks, verify evidence — ksrijith · 2026-10-27
- CoreWeave ARIA agent ships four new live in-chat panels for ML run analysis — _ScottCondron · 2026-10-27
- opentel-mcp brings OpenTelemetry silent-failure detection to Python MCP servers — Thirumalaiboobathi · 2026-10-27
- A hands-on local-model benchmark: porting an old screensaver to Python OpenGL with Qwen 3.8 — prusswan · 2026-10-27
- Hugging Face ships native TPU support via TorchTPU, no torch_xla or JAX needed — ariG23498 · 2026-10-27
- opentel-mcp ships Python support with failure fingerprints matching its JS version — Thirumalaiboobathi · 2026-10-27