HomePath

Inspiration

Housing instability is rarely just a housing problem. It is often connected to financial stress, uncertainty, and a lack of clear information when people need it most.

Whether someone is facing eviction, struggling to pay rent, searching for emergency shelter, dealing with utility shutoffs, or falling behind on mortgage payments, the process of finding help can be overwhelming. Although many government programs, nonprofit organizations, and housing assistance resources exist, they are often scattered across different websites and agencies, making it difficult for people to know where to begin.

We were inspired by a simple idea:

What if people had a trusted guide that could help them understand their situation, identify their next best step, and connect them with the right resources before a housing crisis becomes worse?

We also recognized that people seeking housing assistance are often frustrated, anxious, or overwhelmed. Because of this, we intentionally designed the platform with a calm and welcoming user experience. The soft green color palette, simple navigation, clear progress indicators, and step-by-step guidance were all designed to reduce stress and help users feel supported rather than overwhelmed.

Our goal was to create a platform that not only provides answers, but also provides reassurance and direction during difficult moments.

What It Does

HomePath is an AI-powered housing assistance platform that helps individuals and families navigate housing-related challenges through personalized guidance.

Users can describe their situation in natural language and receive:

Personalized housing guidance
Relevant local assistance programs
Emergency shelter recommendations
Rental assistance resources
Utility assistance programs
Legal aid resources
Mortgage and foreclosure-prevention guidance
A personalized action plan with prioritized tasks

The platform continuously updates recommendations and action plans as users provide additional information, helping them move from uncertainty to action.

Rather than overwhelming users with hundreds of resources, the system focuses on presenting the most relevant next steps based on their specific circumstances.

How We Built It

HomePath was built as a full-stack web application combining conversational AI with structured planning tools.

Frontend

We designed the user experience around clarity and emotional comfort.

The interface includes:

AI-guided conversations
Resource discovery and filtering
Personalized action plans
Progress tracking
Saved resources
Google account authentication

Built with React, Tailwind CSS, React Router, and Google OAuth.

Backend

The backend manages conversations, resources, plans, and user accounts while coordinating AI-generated recommendations.

Built with Node.js and Express.js via REST APIs.

Data Layer

We use PostgreSQL to store:

Table	Contents
`users`	Google-authenticated accounts
`conversations`	Chat history per user
`plans`	Housing action plans
`plan_tasks`	Individual tasks per plan
`saved_resources`	Resources bookmarked by users

AI Layer

The AI analyzes user situations and generates both conversational guidance and structured action plans.

Built with Groq API and Llama 3.3 70B.

The AI pipeline follows a two-stage approach:

Stage 1 — Conversational turn (lightweight, runs on every message):

$$\text{input: message history} \rightarrow \text{output: } {reply,\ situation,\ planAction}$$

Stage 2 — Plan generation (heavy, runs only on explicit user request):

$$\text{input: } {situation,\ conversation} \rightarrow \text{output: } {goal,\ tasks[],\ resources[]}$$

Resource matching uses an urgency-aware scoring function. For a given user situation $s$ and resource $r$:

$$\text{score}(s, r) = \text{base} + \sum_{i} w_i \cdot \mathbf{1}[\text{condition}_i(s, r)]$$

Where:

$\text{base} = 50$
Conditions include category alignment, urgency level, income signals, and concern type (eviction, utilities, legal, mortgage)
$w_i \in {10, 15, 25, 30, 40, 42}$ depending on signal strength

Resources scoring in the range $[90, 99]$ are labeled Strong Match, $[75, 89]$ as Good Match, and $[60, 74]$ as Relevant.

Challenges We Faced

One of the most difficult challenges was balancing flexibility with reliability.

Every housing situation is different. A user facing eviction tomorrow requires a completely different response than a homeowner struggling with mortgage payments or a family seeking affordable housing.

We needed the AI to:

Understand complex situations
Prioritize urgent needs
Recommend appropriate resources
Generate actionable plans
Update those plans as conversations evolved

Another challenge was ensuring the AI could produce structured outputs that the application could reliably use while still maintaining natural and supportive conversations. We solved this by separating the conversational response from the structured data — the model returns a lightweight JSON envelope on every turn, but the full plan draft is only generated when the user explicitly requests it:

$$\text{chat turn cost} \ll \text{plan generation cost}$$

This kept the experience fast and reduced the risk of token-limit failures on the conversational path.

From a design perspective, we also focused heavily on reducing information overload. Housing assistance systems often overwhelm users with large amounts of information. We worked to create a clean experience that highlights only the most relevant actions, resources, and next steps.

What We Learned

This project taught us that successful AI applications are not just about generating answers.

We learned how to:

Design AI-driven user experiences
Build structured AI workflows
Create reliable prompt architectures that separate tone rules from output schemas
Integrate conversational AI into production systems with graceful degradation
Develop scalable full-stack applications
Design technology with empathy and accessibility in mind

Most importantly, we learned that AI can be most impactful when it helps people navigate high-stress situations and take meaningful action.

Accomplishments We're Proud Of

Built a complete AI-powered housing guidance platform
Created personalized housing action plans that update as circumstances change
Integrated dynamic, urgency-aware resource recommendations
Added real-time progress tracking and plan management
Designed a calming, user-centered experience
Supported multiple housing scenarios: renting, emergency housing, utilities, and mortgage challenges
Built a model fallback chain that degrades gracefully when primary models are rate-limited
Implemented per-user data isolation with Google OAuth and PostgreSQL

What's Next

HomePath currently focuses on helping users navigate housing challenges through personalized guidance and curated resources in Atlanta.

Future improvements include:

Expanding resource coverage beyond Atlanta
Real-time government and nonprofit resource verification
Document analysis for eviction notices, mortgage statements, and housing paperwork
Multilingual support
Appointment and deadline reminders
Partnerships with housing assistance organizations
Vector similarity search for more precise resource matching:

$$\text{score}(s, r) = \frac{\vec{v}_s \cdot \vec{v}_r}{|\vec{v}_s| |\vec{v}_r|}$$

Where $\vec{v}_s$ and $\vec{v}_r$ are embedding vectors of the user situation and resource description respectively.

Our vision is to make housing assistance easier to understand, easier to access, and less stressful to navigate — helping people find stability before a crisis becomes irreversible.