FertilitySync

Poster

Overview Build an AI-powered real time emotional support system designed for individuals undergoing fertility treatments. It delivers real-time mood tracking, 24/7 support, alerts support system if needed and provides clinician-vetted guidance to improve mental wellbeing and, in turn, strengthen fertility outcomes. Since it is a long and iterative process, personalized support is necessary for individuals going through their fertility journey.

Objectives -Instant and 24*7 instant emotional support analysing real time mood. -Clinically aligned safety net and detecting red flag moments like suicidal thoughts and escalating appropriately -Improve fertility rates, reducing stress, anxiety and emotional volatility, which are proven to affect treatment outcomes -Strengthen relationship harmony by preventing emotional overload, guilt, reducing partner conflict and lowering the risk of breakdowns and separation -Reduces unnecessary IVF cycle repetition due to mental health issues, reduces emotional burnout, and supporting continuity through difficult phases even there’s no one else to support you in real life. -Supports NHS by reducing mental health loads, improving early triage and providing evidence based self management pathways aligned with clinicians. -Delivers a scalable, culturally adaptable support framework that works for individuals from every background.

Explored all 3 tracks -A+B+C

Technical Details- Used Holistic AI to build an AI agent for this use case Used Valyu API to gather data and test cases for the dataset

For Track A Built robust AI agent framework to understand emotions from the chat and provide emotional support approved by doctor Built framework to use different LLM models like Claude, OpenAI,and DeepSeek The agent identifies the user’s mood and recommends if external intervention is necessary based on the conversation

For Track B Have a score to justify mood classification Identify keywords from the prompt for emotional classification Sticks to the domain for the conversation Obtained LangSmith UI traces for observability Built dynamic tags to analyse runs Generated metrics such as the number of scoring requests, cost and cache hit rate Used metadata to analyse runs

Track C Diagnoses if injections are included Carried out exploratory red teaming to evaluate bias risks in the responses of the agent Executed attack approaches (pretend and role-play) to test the robustness of the agent Implemented prompt leaking and token bombing Learnings This is the first time the entire team is learning about AI agents and we used it to build a personal mental health companion. We all learnt a lot about AI agents,NLP, LangChain and LangSmith this weekend