▶ Watch ↗1:50:32
Conference affiliation: OpenAI
Oliver Steffmann develops technical approaches to financial work, from testing algorithmic trading strategies to connecting enterprise AI assistants with company data. An applied AI engineer at OpenAI at AI Engineer NYC 2026, he brought experience in asset management, investment banking and cloud architecture to a practical question: how can financial institutions adapt AI to their existing methods and systems?
Oliver began his career in his own startup in Germany. He later headed municipal trading technology at a major investment bank in New York and worked in architecture, software-development management and consulting for international financial institutions. His AWS author biography describes that background; his subsequent work at AWS included solutions architect and principal solutions architect roles.
His 2021 algorithmic trading implementation connected Amazon SageMaker with Backtrader, a framework for testing strategies against historical data. The model addressed a specific trading outcome: whether a long or short position would reach a profit target before being stopped out. Oliver trained the classifier on one portion of the historical data and tested the strategy on the remainder.
The implementation also covered the engineering around the prediction: storing market data and results, packaging strategies in containers, and measuring profit and loss and risk-adjusted performance. The accompanying Python notebooks and infrastructure supplied an educational implementation, without establishing that historical results would translate into future profitability.
Oliver’s later AWS writing addressed how organizations introduce AI assistants into everyday work. In August 2025, he, Krishna Pramod and Mo Naqvi proposed an incremental approach to deploying Amazon Q Business: choose a bounded workflow, connect supported data sources, test answers against ground truth, and expand with feedback from participating teams. Access controls and measurement were part of the rollout, alongside adoption by an initial group of users.
In January 2026, Oliver and five colleagues developed a contract-management reference architecture combining Amazon Quick Suite with Bedrock AgentCore. A coordinating agent routed contracts to specialists examining legal obligations, financial and operational risk, and regulatory compliance, then consolidated their findings. The collaborative design connected an employee workspace to specialized processing, document stores and structured data.
At AI Engineer NYC, Oliver and Stefano Fabbri explored a related problem through a financial-adviser workshop. An adviser needed to explain changes in a client’s portfolio and prepare a presentation about possible rebalancing. Market data and filings could help explain events, but calculating which holdings contributed to a gain or loss required the firm’s own portfolio records.
The workshop distinguished three ways developers could adapt an existing AI workspace: a reusable skill could capture the firm’s reporting method; a plugin could connect that method to internal records through Model Context Protocol tools; and a plugin extension could customize the interface used for the task. This gave a concrete alternative to building a separate application when employees already worked in an approved ChatGPT or Codex environment. It still required developers to manage authentication and secure the connected tools.
Oliver and Vik Bomma’s October 2026 guide to failover across OpenAI model providers examines what happens when an operation fails or times out. Its central distinction is between having another provider available and knowing whether a particular workflow can safely move to it.
A read-only fraud assessment may be repeated after a timeout if its answer has not been used and enough time remains. An agent that may already have taken action presents a different problem: the system must establish what happened and prevent a retry from duplicating the action. Switching providers does not resolve that uncertainty.
The guide also rejects model-name matching as sufficient evidence of interchangeable behavior. An alternative route needs its own evaluations, capacity checks and approved data handling. Recovery therefore depends on the operation being performed, the state it may have changed, and the behavior of the replacement route.
▶ Watch ↗1:50:32
Loading more talks and highlights…