Jonny Press
Over the last few months we’ve been building AI accelerators aimed squarely at the kdb+ platforms we know best. We now have three demos we’re showing to clients. Each one tackles a real Capital Markets problem, sitting on top of the domain expertise and KX engineering skills Data Intellect has built up over the years.
We have learned a lot along the way. Agentic systems are not a free lunch. You need to think carefully about which agent does what, which tools to build, how much context and data to supply, how to keep costs under control, and how to capture every useful piece of data along the way (which aligns neatly with what kdb+ is already good at). Live demos in a probabilistic world also come with crossed fingers every single time. Each demo is a trust exercise, and some of the engineering work is more about reducing the number of fingers you have to cross.
The other thing we’ve learned is that the AI part is rarely the hard part. The hard parts are the data plumbing and the domain understanding. And most parts of it shouldn’t be AI- rules based determinism is preferable as often as possible.
Surveillance Analyst
All good trade surveillance tools provide the ability to investigate and action alerts through a set of pre-configured screens. These serve the majority case- this chat-based investigative tool comes into play for both extended investigations where access to underlying data is required, as well as higher level summary analysis, such as “why is there an increased volume of alerts today?”.
An analyst can leverate the power of KDB-X to dig across structured alert and trade data and unstructured context- communications, news, internal documents- through natural language, and the agent understands enough of the domain to ask the right follow-up questions. The agent can be customised to the requirements of the individual analyst, utilising specific regulatory handbooks and internal policy documentation.
Operations Analyst
A multi-agent system that diagnoses issues inside a kdb+ environment and recommends fixes. Automatic classification and investigation of issues as well as chat based initiation, with cost aware orchestration. Investigation and decision traces are tracked for refinement, training and audit purposes.
An example of a common question our support teams face is “why is my query slow?”- this could be caused by any of system degredation, resource consumption, code errors, data issues or user contention. There’s a lot of separate, asynchronous timeseries data to reason across to get the correct answer to that question.
AI Trader Backtesting
A backtesting framework to compare and refine trading strategies, with LLM inference running under NVIDIA NEMO on NVIDIA GPUs. kdb+ handles the time-series heavy lifting, feeding the data and capturing the reponses of the trading model running on the GPU.
We’d like to show you any of the three accelerators. We are happy to share our experience and findings, and we’d love the feedback. If you run a kdb+ or KDB-X platform and you’re trying to work out where AI fits, we’ll show you some of the parts we think it works for.
Book A Demo
Share this:
