HerVoice

Inspiration HerVoice was inspired by the real challenges women in STEM face, feeling unheard in meetings, navigating biased feedback, lacking mentorship, and managing burnout in male-dominated spaces. We envisioned an AI-powered support assistant that listens, understands, empowers, and guides women through these challenges with empathy and actionable insights.

What it does HerVoice is a conversational AI assistant designed to support women in STEM by providing:

• Emotional validation and motivation

• Career guidance and mentorship suggestions

• Safe responses to bias and exclusion

• Real-time answer verification and citation

• Anonymous, stigma-free access to support

It integrates multi-query Retrieval-Augmented Generation (RAG), hallucination grading, and query rewriting, all wrapped in a friendly, empowering Streamlit interface.

How we built it We built HerVoice using:

• Google Gemini as the core LLM for empathetic dialogue

• LangGraph for agentic RAG workflows with fallback and retry logic

• PGVector to store and retrieve curated books, toolkits, and handbooks

• Tavily for up-to-date web search augmentation

• Grading modules for hallucination and relevance checks

• Streamlit for a clean, supportive user interface

• Each component is modular, allowing personalized feedback grounded in real-world resources.

Challenges we ran into

• Balancing tone: ensuring the assistant felt empathetic without being overly generic

• Avoiding hallucinations in sensitive mentorship advice

• Integrating citation-aware RAG with dynamic fallback to web search

• Designing scalable conversation test cases grounded in lived experiences

Accomplishments that we're proud of

• Created a realistic and relatable test case suite using 60+ real-life queries from women in STEM

• Developed an end-to-end agentic RAG pipeline with hallucination filtering

• Indexed trusted handbooks (like Lean In, AAUW, UNESCO, and NCWIT) into a vector store for evidence-based support

• Delivered a visually appealing Streamlit interface that feels welcoming and safe

What we learned Technical: Implementing MMR+RRF in multi-query RAG dramatically improves relevance

Product: Bias and burnout show up subtly in user language, recognizing this helped shape more effective prompts

Human: Small phrases like “I don’t feel heard” can signal deeper workplace challenges. Listening closely is everything.

What's next for HerVoice

• Integrate anonymous reporting templates for users navigating hostile environments

• Partner with women-in-tech networks to expand HerVoice’s reach and training data

• Launch mobile and Slack-compatible versions for in-context support

Built With

• and-pandas-for-test-case-evaluation.-github-handled-version-control
• and-relevance-scoring-to-ensure-users-get-trustworthy
• and-the-system-was-designed-to-be-cloud-deployable-using-platforms-like-streamlit-cloud-or-google-cloud.-the-tool-leverages-a-modular-architecture-with-hallucination-grading
• and-uses-pgvector-with-postgresql-to-store-and-retrieve-embeddings-from-curated-dei-handbooks-and-toolkits.-we-used-tavily-for-web-search-fallback
• context-rich
• hervoice-was-built-using-python-and-streamlit-for-a-clean
• openai-or-sentencetransformers-for-embeddings
• query-rewriting
• real-time-conversations
• supportive-frontend-interface
• with-langchain-and-langgraph-powering-its-agentic-workflow-and-multi-query-rag-system.-it-integrates-google-gemini-via-the-generative-ai-api-for-empathetic