Opetek has launched ARIUS, a quantitative reasoning platform for institutional trading desks, which is being piloted by traders and quants at a tier-one global investment bank across live FX options workflows in New York, London and Singapore. A second tier-one institution has been approved and onboarded, with go-live expected in the coming weeks.
Designed for front-office desks, ARIUS is delivered as a standalone application sitting above existing data, analytics and model layers, connecting market data, pricing models, analytics and news through a natural-language interface so that traders and quants can generate bespoke analysis and code without building one-off tooling each time. Opetek describes the problem it addresses as four sources of decision friction: fragmented workflows across market data, analytics, risk, research and communications systems; cognitive overload from data volumes; delay introduced by manual coding and bespoke analysis; and undocumented assumptions that leave decisions difficult to inspect, reproduce and monitor.Hallucination in numerical work is the constraint the architecture is built around.
“We don’t just give the trader the raw output of a large language model,” Varqa Abyaneh, Founder and CEO of Opetek, tells Market & Alt Data Insight. “We use the model to generate a program that captures the verbal and numerical reasoning. The trader sees a clear textual answer with the relevant tables and calculations, but underneath it sits deterministic code that can be re-run independently without calling the LLM again, inspected and debugged, ensuring the outputs are reproducible.”
The company says the assumptions, methodologies and market data inputs behind an output can all be inspected and challenged by traders and quants.
Opetek runs on a private cloud using standard commercial large language models, with its own multi-agent layer above them rather than a proprietary or retrained model. Model risk governance at large banks is built around models whose behaviour can be documented and re-tested.
“Banks are understandably cautious about models with bespoke training because they require additional governance and validation, which can slow onboarding,” says Abyaneh. “It is generally quicker to deploy using a standard, approved model and differentiate through the architecture around it. Bespoke training may improve an answer, but additional work needs to be done to ensure the improvement is consistent, material and safe.”
Market data comes from a partnership with LSEG (Refinitiv), alongside web access, with clients able to connect their own data, chat records and pricing functions. Connectivity is available through API or MCP. Opetek reports that its method of connecting reduces token costs by around 15 times compared with routing query results through a context window, though it has not published a methodology for the figure and Abyaneh declines to describe the mechanism.Use cases the company has put in front of the pilot desk include testing whether market reactions to political statements have become damped over time, and classifying volatility regime shifts to inform market making. Abyaneh says the firm is building capability up in stages, establishing that the system can reason on market data and justify its answers before adding analytic functions, with a test harness run against each release.
Prior to founding Opetek, Abyaneh was previously Chief Product Officer at Quantile, the compression and portfolio optimisation firm acquired by LSEG in 2022, and before that in derivatives trading, quantitative analysis and product roles at banks including HSBC.
FX options is the first asset-class application of the underlying engine, with equities, fixed income, commodities, credit and digital assets named as candidates for extension.
Subscribe to our newsletter
