Case study —
Sophie LLM
( Overview )
Sophie Large Language Model (LLM) is a dashboard that hosts a variety of AI-powered tools, providing users with the ability to interact with diverse datasets for various use cases.
( Scope )
User interface design of dashboard and tools for various use cases. Involved in the development and execution of a proof-of-concept to launch to internal staff.
( Role )
Role: Product Designer
June 2023 - Ongoing
The Dashboard
The Sophie dashboard aims to offer a combination of general LLM solutions, encompassing Q&A, translation, and summarisation capabilities. It addresses specific user cases, ensuring responses are not only more accurate but also tailored to the unique needs of specific user groups.
Since its introduction, Sophie has garnered significant attention from internal staff. Its ability to address pain points and solicit a wider range of use cases has contributed to its growing popularity.
( Use Cases )
Market & Investment Insights
A database of all market and investment analysis and recommendations for users to keep up-to-date with the latest market trends and economic changes.
( Background )
The bank has a research team that writes and publishes investment content for clients and staff. Investment writers generate talking points and insights to aid relationship managers to better communicate with their clients.
( Problem )
It is time consuming to source through the research bank to find suitable insights for specific topics and conversational points. It is also difficult to consolidate research across time horizons to find trends.
( How Might We )
How might we provide relationship managers with an easy access to all market and investment analysis and recommendations for them to keep up to date with the latest market trends and economic changes so as to better facilitate the way they advise their clients.
( Solution )
Large Language Model (LLM) chatbot backed by a database of all research, market and investment analysis
Query the data universe
Users can ask questions pertaining to the universe of research, market and investment data. An answer will by formulated by the model using the latest and most relevant data in the database.
Users can toggle the sources to view the original documents where the data chunks are being extracted from.
Users can toggle the sources to view the original documents where the data chunks are being extracted from.
Translation
Users can translate the answers by simply choosing the desired output language from the dropdown.
Joice
Voice-to-text application utilising the Whisper model to to automate any manual process of transcribing and translating any recorded audio, allowing users to easily analyse the textual output.
( Background )
Whisper, a general-purpose speech recognition model trained on a diverse audio dataset, is capable of multitasking, including multilingual speech recognition, translation, and language identification.
One use case of voice-to-text technology is in accelerating and automating manual processes, notably for Compliance Risk Officers. These officers review recorded calls, or "voice logs," between Relationship Managers (RMs) and clients to perform audit, analysis and reviews to ensure regulatory compliance.
One use case of voice-to-text technology is in accelerating and automating manual processes, notably for Compliance Risk Officers. These officers review recorded calls, or "voice logs," between Relationship Managers (RMs) and clients to perform audit, analysis and reviews to ensure regulatory compliance.
( Problem )
Internal staff often need to transcribe voice recordings manually and the process is long and tedious. On average, transcribing a voice recording manually can take approximately 4-6 times the duration of the recording itself.
Voice logs can also come in foreign languages and translators are necessary for understanding foreign conversations, but hiring them incurs costs and takes months to be translated due to the requirement of hiring a translator. If there is no translator available, these voice calls can't be translated and sit in the backlog.
Voice logs can also come in foreign languages and translators are necessary for understanding foreign conversations, but hiring them incurs costs and takes months to be translated due to the requirement of hiring a translator. If there is no translator available, these voice calls can't be translated and sit in the backlog.
( How Might We )
How might we provide compliance risk officers with an all-in-one voice-to-text tool so as to cut down the manual process of transcribing voice logs and reducing the dependency on translators.
( Solution )
Voice-to-text tool for effective transcription and translation of voice logs
95%
time and cost savings
Transcribing a one-hour recording manually takes 4-6 hours while Joice can take as little as 5-10 mins. With greater time saved, staff can instead focus on the next step and actions generated from the transcription.
Transcribing a one-hour recording manually takes 4-6 hours while Joice can take as little as 5-10 mins. With greater time saved, staff can instead focus on the next step and actions generated from the transcription.
Upload audio files
Users can simply upload single or multiple audio files and select the suitable number of speakers for the tool to identify and annotate respective speakers.
Transcribe
English transcription of the audio file will begin with the speakers and time stamps being annotated.Users can either download or copy the transcription output once completed.