About PathBridge AI
PathBridge AI is a human-in-the-loop, agentic AI prototype designed to help survivor-support organizations create safer and more realistic economic recovery pathways for people who have experienced exploitation. Our motto is: AI-guided recovery, human-led decisions.
Inspiration
The main inspiration behind PathBridge AI was the possibility of building something that could have real-world applications. We wanted to design an AI system that was not just technically interesting, but meaningful in a human safety context. Survivors of exploitation often face complex barriers when trying to rebuild economic stability, such as limited work history, unsafe transportation, trauma-related constraints, documentation issues, childcare responsibilities, housing instability, and urgent income needs.
A normal career recommendation system may suggest the highest-paying or fastest available job, but that does not mean the pathway is safe, realistic, or survivor-centered. This inspired us to think about how AI could be used more responsibly: not as a replacement for human support, but as a tool that becomes more reliable through human guidance.
Our idea is based on the belief that AI can be useful in sensitive real-world problems if it is designed with the right boundaries. PathBridge AI keeps human advocates in control while using AI to organize information, identify possible pathways, flag risks, and explain tradeoffs.
Instead of asking only, “What job fits this person?”, our system considers safety, existing skills, personal constraints, timeline, and human advocate review together before suggesting a recovery pathway.
What We Built
PathBridge AI has two separate interfaces:
User Support Request Interface A user or support worker can describe the situation in natural language. The user does not need to fill out a complex form or understand workforce databases.
Human Advocate Review Workspace This is where the real decision-support work happens. The system uses LLM-powered agents to extract needs and constraints, retrieve relevant public-data records, compare recovery pathways, flag safety concerns, and support discussion with a human advocate.
The system uses a hybrid architecture:
i. LLM-based agents for natural-language understanding, case summarization, pathway explanation, advocate discussion, and final plan generation.
ii. Rule-based guardrails for safety checks, privacy protection, scoring, and human-review control.
iii. Public-data retrieval from source-labeled workforce and survivor-care datasets inspired by O*NET, BLS/OEWS, CareerOneStop, and public survivor-support reports.
The final recovery plan is generated only after the human advocate reviews, modifies, rejects, or escalates the AI-generated pathways.
How We Built It
We built the prototype using Python, Streamlit, and the OpenAI API. Streamlit allowed us to quickly create a low-code interactive interface with separate user and advocate views. The OpenAI API powers the LLM agents used for intake extraction, advocate analysis, discussion, and final plan generation.
The system follows this workflow:
User describes situation
↓
LLM Intake Agent extracts needs and constraints
↓
Public-data retrieval layer finds relevant pathway records
↓
Rule-based safety and feasibility checks score pathways
↓
LLM Advocate Analysis Agent explains options
↓
Human advocate reviews, modifies, rejects, or escalates
↓
LLM Final Plan Agent generates reviewed recovery plan
What We Learned
We learned that responsible AI for human safety requires more than generating good answers. It requires careful system design. In this problem space, AI should not act as an autonomous decision-maker. It should act as a structured assistant that helps human advocates see options, risks, and tradeoffs more clearly.
We also learned that human guidance can make AI systems more reliable. The AI can organize complex information and suggest possible directions, but a human advocate can correct, modify, reject, or contextualize those suggestions. This feedback loop is especially important when the problem involves vulnerable populations.
Another major lesson was that hybrid AI systems are useful for sensitive domains. LLMs are strong at understanding messy natural language and explaining complex tradeoffs, while rule-based components are better for predictable safety checks, privacy controls, and human-review requirements.
Challenges We Faced
One challenge was deciding how much autonomy the AI should have. Since the project involves vulnerable populations, we intentionally limited the AI’s authority. The system does not directly tell a survivor what to do. Instead, it sends all detailed planning to the human advocate review workspace.
Another challenge was data availability. Real survivor-level data is sensitive and should not be used casually. To address this, we avoided private survivor data and used public-data-style records for workforce pathways, wages, training options, and survivor-care context.
We also had to design the interface carefully. We separated the user-facing request screen from the advocate review workspace so that sensitive analysis, risk flags, and decision-making remain under human supervision.
Impact
PathBridge AI helps survivor-support organizations turn complex situations into structured, safer recovery plans. Instead of manually searching disconnected resources, advocates can use the system to identify possible pathways, compare risks, discuss alternatives, and generate a reviewed plan.
The most important design choice is that PathBridge AI keeps humans in control. The AI can summarize, retrieve, compare, and explain, but the survivor and trained advocate remain the final decision-makers. This makes the system more realistic for real-world use, where trust, safety, and human judgment matter as much as technical performance.
Log in or sign up for Devpost to join the conversation.