A global banking major set out to build something its traders had never had before: a truly LLM-integrated chatbot for conversational GTT support. Up to that point, traders relied on a traditional portal that demanded extensive manual searching just to retrieve GTT-related data. To speed things up, the bank had introduced the Optimus chatbot, but that earlier version leaned entirely on predefined responses and adaptive cards, which meant users still had to repeatedly search for values through dynamic typeaheads, reload responses after any change, and navigate a bulky, inflexible UI. The experience was slow, unintuitive, and heavy, exactly the opposite of what a conversational tool should feel like. Moving toward genuine LLM-driven conversation opened up new challenges too: NLP misinterpretations, potential data confidentiality risks, and stringent InfoSec approvals all had to be worked through. Getting this right meant ensuring accurate intent detection, predictable responses, and secure data handling from day one, not as an afterthought.
Ascendion’s GTT team pioneered the bank’s first LLM-integrated chatbot using CALM, LLAMA 70B, and RASA, building a conversational assistant that interprets simple English questions, identifies the data users need through LLM-based intent detection, and retrieves results cleanly, without the constraints of adaptive cards or repetitive value selection.
The team built and refined a local prototype, iterating through aggressive testing cycles and collaborating closely with RASA specialists to rigorously train the model against drifting into unrelated topics. Python powered the backend actions, with Bitbucket handling CI/CD, ensuring the chatbot’s behavior remained controlled, relevant, and high-performing throughout. The result is a fast, minimalistic, and highly responsive experience that fundamentally changes how traders interact with GTT data.
Tech Stack:
Completed the project in record time of approximately 3 months, ready for user demo.
Became the bank's first Reference Data team to deliver an LLM-enabled chatbot.
Delivered significant time savings compared to the legacy portal and earlier Optimus chatbot.
Improved performance, eliminating the need for repeated manual searches.
Delivered a cleaner, more minimalistic, user-friendly interaction.
Sparked interest across multiple teams now planning their own LLM-driven chatbot use cases.
