Timeline:
8 Weeks
Role:
AI Conversation Designer
Figma | VS Code | Github | Telegram | Vapi
Case Summary
Problem
Families in food deserts depend on government or non-profit food programs that may carry inconsistent rules, calendars, and paperwork. Support is hard to discover, breaks down seasonally and geographically, and demands more time and travel than many households can spare. Because each organization collects its own personal data, providers can't coordinate, and families face real privacy risks, leaving communities reacting to hunger rather than preventing it.
Context
A class Sponsored Studio with SAS, a software company that specializes in data management, advanced analytics, and artificial intelligence (AI).
Four-person team proof of concept for a coordinated SNAP food-access services: a phone line and a chat assistant for recipients, a kiosk for ordering SNAP-eligible food, and a portal where pantries analyze inventory. I owned the two conversational agents.
Outcome:
Two working prototypes:
Chat agent powered by Telegram for UI, Gemini 2.5 Flash, Python, and VS Code
Voice Agent powered by VAPI (Orchestrator)
Approach to Project
Designing for the user's worst possible day.
Leveraging AI tools (Claude) to augment my Python coding skill
Design Challenge
How might we support families in food deserts in finding and understanding the food resources they qualify for, while keeping families in control of their personal information?
Relative Competitive Analysis
• Existing tools are dispersed across use cases, either serving multiple domains or a
specific need.
• Some tools have eligibility information that may be difficult to understand
• An opportunity to create a universal platform for understanding SNAP rules and guidance to alternative food aid resources
What the stats say:
Food insecurity affects a large share of U.S. households with children
Affordability and storage barriers predict food insecurity
Persona Profile
Creating a persona profile keeps me grounded in who I am designing for. This persona represents a working parent with children who has just begun using SNAP benefits and is looking to maintain food security in her household.
Designing a Chat Agent Flow
Before coding the chat agent, I first created a flow for how it could behave and what users might ask. At the time, I named the chat agent "Cura" but received feedback that it should remain brand-less. Additionally, I was beginning to frame how the chat agent should format its responses and keep the conversation naturally flowing. I understood there could be technical limitations in the chat agent's output, so I considered realistic outputs.
Working on the System Prompt
To ensure the SNAP chat agent acted appropriately and understood its role, I created a markdown file including it's persona, the tools the model allows, and what the chat agent's delivery and tone should be. This meant emphasizing that it should write at a 6th-grade reading level and avoid unnecessary jargon.
From VS Code to Telegram
With the assistance of claude and my begginer knowledge of python, I created a codespace where the chat agent could operate and leverage Telegram to host my UI.
Creating the Voice Agent
To ensure this agent was accessible across various mediums, I was curious on how a voice agent could behave and it's potential use case. The biggest motivator behind creating the voice agent was in emergency scenarios where recipients do not have time to text large paragraphs about their situation.
Voice Agent Flow
This was one of the flows I drafted to understand how the voice agent can respond to certain recipient inquiries. From feedback from the team and SAS advisors, the menu script may be harmful in emergencies.
Developing the System Prompt
One key difference I found when developing the voice agent system prompt was emphasizing how it should handle emergency scenarios. I focused more on its tone and pace. I also explored outbound call opportunities where the voice agent can call recipients to remind them about appointments or balances.
Leveraging VAPI
VAPI acted as an orchestration layer, saving me hours of coding. I utilized it to develop the cost-effective nature of the agent, paste the system prompt, and gather files that acted as the source of truth.
Next Steps
This project expanded my knowledge and curiosity about how chat and voice agents can be implemented in everyday life, such as the public government support domain. Given more time, I would strive to implement multilingual support, add generative UI so recipients can complete paperwork straight from the chat environment, and expand the guardrails for how the agents behave.