Objectives
The aim is to develop a proof-of-concept chatbot that can handle open-ended user inquiries using Girl Effect content developed by a health and gender team. This content will draw from years of Q&A sessions with real girls in our target markets. This content, combined with data on local services such as health clinics or mental health services, will help equip young people, when ready, to access these resources, ensuring their physical and emotional needs are met.
Within this piece of work, we’ll set up the foundational infrastructure to support and possibly deploy the chatbot. It’s important to note that scaling this infrastructure and releasing the chatbot to the public aren’t included in this phase. Instead, we’ll rely on internal users and focus groups for evaluation and testing to validate our proof of concept.
Key Deadlines
We will be taking a phased approach to delivering conversational AI. This TOR defines the first Phase to be delivered within 6 months of the project start.
Within this 6 month deliverable, we will focus on a proof of concept to ensure that we can provide a relevant, safe and engaging conversation chatbot. For this phase, we have identified three key learning questions that will inform our understanding of the viability of available vendor and open source models to perform the specific capabilities to the performance standards required by Girl Effect.
Tone, language and Relevance: Can the latest technologies in NLP and Large Language Models deliver a highly personalised experience in tonality, language and relevance and enhance personalised experience by directly providing users with the content they request in a tone and language that meets their cultural preferences?
Safeguarding: Can the latest technologies continue to provide a safe environment for our users and enhance our ability to detect sensitive disclosures and direct users to resources and services when they need them?
User Engagement: Will integrating Gen AI and switching from a menu-based experience to a conversationally-led experience enhance our user engagement?
Deliverables
You will deliver a proof-of-concept chatbot that can handle open-ended user inquiries using Girl Effect-approved content, vetted GenAI models, and chatbot and infrastructure frameworks.
We aim to rigorously evaluate various open-source and proprietary language models. Our research methodology will be meticulously crafted to address each learning objective, fulfill all requirements, and encompass the entire project scope.
We will handpick models for testing based on their proven efficacy, general recognition in the field, and substantiated outcomes. An initial set of models will be established at the project’s outset, anticipating iterative refinements over the six months, influenced by emerging insights and model availability.
The vendor will employ an evidence-based testing framework to assess each model’s Accuracy, Efficiency, and Cohesiveness, considering factors like tone, language adaptability, and contextual relevance. We expect the vendor to implement a suite of automated tests, enabling precise and consistent comparative analysis of each model’s performance.
Your role also entails delivering expert recommendations for model hosting, maintenance infrastructure, chatbot frameworks, and the necessary supporting infrastructure. You are to utilize evidence-based methods for all evaluations. You will be accountable for identifying and detailing all associated costs for each recommended framework and infrastructure solution, whether Cloud, SaaS, or other related expenses.
Requirements
Chatbot Framework: It’s essential to evaluate the chatbot platform and framework to choose the best-fit solution. You will be responsible for accessing and selecting the appropriate framework for our next-generation conversational GenAI chatbots.
Vetted Content: Ensuring our chatbots disseminate exclusively verified information is crucial. Your task is to build a chatbot infrastructure that guarantees user responses are confined to content sanctioned by Girl Effect, encompassing approved knowledge and service-related data. Employ and rigorously test suitable methodologies to ensure the model delivers safe and vetted content consistently.
Technical safeguarding: The infrastructure must incorporate mechanisms for identifying safeguarding disclosures and initiating the correct response protocols. The proof of concept (POC) needs to showcase the ability to recognize safeguarding instances and seamlessly integrate them into the appropriate action-oriented workflows.
Mix-code Languages: Girl Effect operates in areas where low-resource languages prevail. (languages that lack large monolingual or manually crafted linguistic resources sufficient for building statistical NLP applications). Not only are users in these regions speaking in low-resource languages, they also often mix languages and local slang, speaking in what is termed mix-coded languages. Relevant techniques should be employed and evaluated to facilitate conversations in specific mixed-code languages. Specifically in scope will be Hinglish for the Indian market and Sheng for Kenya.
Culture and Tone: The design of Girl Effect’s chatbot user experience must mirror the local cultural context and resonate with the user’s tone. Employ and evaluate effective techniques to capture the youthful essence that aligns with the Girl Effect brand identity.
Speech to Text and Text to Speech: Understanding the capability of speech-to-text and text-to-speech chatbots for integration into IVR or other lo-fi systems for processing and communication through voice notes. This scope should include testing speech-to-text and text-to-speech for the specified low-resource and mix-coded languages.
Production infrastructure: Based on the POCs implemented in this workstream, a production infrastructure should be created capable of supporting and testing the relevant models and chatbot infrastructure.
Apply via :
boards.greenhouse.io
