Last Mile Navigator — Project Story
Inspiration
The idea started with a simple observation: UMass Amherst already has the resources students need — they just can't find them when it matters most.
The Dean of Students Office has distributed \$285,656 in emergency microgrants to 605 students since Fall 2022. The Campus Pantry at 472 North Pleasant Street provides roughly 10 days' worth of free groceries per visit. CCPH offers 24/7 crisis support at 413-545-4357. These are real, funded, functioning programs.
Yet the utilization gap is staggering. An internal UMass study found that approximately 25\% of students experience food insecurity. If we model the affected population as:
$$N_{\text{affected}} \approx 0.25 \times N_{\text{students}}$$
For a student body of $\sim$30,000, that's roughly 7,500 students — but the Campus Pantry, even in its successful first year, serves a fraction of that number. The gap isn't supply. It's navigation.
We realized this is a last-mile problem, borrowing the concept from logistics: the most expensive and failure-prone part of any delivery system is the final stretch to the end user. Universities have built the warehouse (resources), the trucks (staff and offices), and the roads (websites and handbooks). What's missing is the last mile — getting the right resource to the right student at the right moment, in a way they can actually act on.
What We Learned
1. Cognitive load under stress is the real barrier. When a student is food-insecure, behind on rent, and failing a class simultaneously, their available cognitive bandwidth for navigating bureaucracy approaches zero. The relationship can be loosely modeled as:
$$P(\text{access}) \propto \frac{1}{1 + e^{-k(\text{bandwidth} - \theta)}}$$
where $\theta$ represents the minimum cognitive threshold needed to navigate the application process. As stress increases and bandwidth shrinks, the probability of accessing help drops sigmoidally — not linearly. This means the students who need help most are mathematically the least likely to find it without intervention.
2. The problem is multi-dimensional, not single-issue. Students rarely face one crisis in isolation. Job loss cascades into food insecurity, which cascades into poor academic performance, which cascades into mental health deterioration. A useful agent must detect and address the vector of needs, not just the one the student explicitly mentions.
3. Social capital is an invisible variable. First-generation students and international students consistently underuse support services — not because they need them less, but because they lack the institutional knowledge that comes from having college-educated parents, upperclassman mentors, or cultural familiarity with American university systems.
How We Built It
We built the Last Mile Navigator on the UMass GenAI Platform (genai.umass.edu) using the Agent Builder feature.
Architecture
The agent has two components:
System Prompt (Instructions): A ~2,000-word structured instruction set that defines six core behaviors — need detection, urgency prioritization, warm tone, communication drafting, proactive inference, and safety guardrails. The prompt includes a trigger-word mapping table that connects student language (e.g., "can't afford," "skipping meals") to specific resources.
Knowledge Base (File Context): A comprehensive directory of UMass Amherst student resources, sourced entirely from official university websites. This includes eligibility criteria, contact information, physical locations, hours of operation, and application processes for programs like SCERF Microgrants, Short-Term Loans, the Angel Fund, the Campus Pantry, CCPH, Disability Services, Student Legal Services, and more.
Model Selection
We chose GPT-5 (Azure OpenAI) with temperature set to 0.7 — lower than the default 1.0 to reduce hallucination risk while preserving conversational warmth. A support agent that invents a nonexistent grant program or gives the wrong phone number could cause real harm.
Design Principles
- Empathy over efficiency: The agent says "That sounds really stressful" before it says "Here are your options."
- Steps over lists: Instead of dumping 10 links, it identifies the single most urgent action and walks the student through it.
- Action over information: It doesn't just tell you about the microgrant — it offers to draft your hardship statement.
- Safety first: Any mention of self-harm or crisis immediately surfaces CCPH's 24/7 line and the 988 Lifeline. The agent never attempts to counsel.
Challenges We Faced
1. Balancing specificity with accuracy. The more specific the agent's responses (exact dollar amounts, exact eligibility rules), the more useful they are — but also the more likely to become outdated. We addressed this by sourcing all data from official UMass websites as of April 2026, and by having the agent direct students to contact offices for final confirmation rather than guaranteeing outcomes.
2. Tone calibration. Early prompt iterations were either too clinical ("Based on your stated circumstances, the following resources may be applicable...") or too casual ("Hey! Don't worry, we got you!"). Finding the right register — warm, respectful, professional but human — required multiple rounds of testing with different crisis scenarios.
3. Handling multi-need interactions. A student saying "I lost my job and can't focus on classes" has at least three needs: financial, food, and academic. The agent must prioritize without overwhelming. We solved this with a six-tier urgency hierarchy (Safety > Basic Needs > Financial > Academic > Mental Health > General Navigation) embedded in the system prompt.
4. The empathy-accuracy tradeoff. An empathetic agent that gives wrong information is worse than a cold one that gives correct information. We resolved this by keeping temperature at 0.7, grounding all responses in the uploaded knowledge base, and instructing the agent to say "I'm not sure about that specific detail — here's who to contact" rather than guessing.
"The information Maria needed already existed on campus. The emergency grant was on page 47 of the student handbook. The food pantry was mentioned in an orientation email she deleted in August. The agent didn't create new resources — it closed the last mile between a student in crisis and the help that was already there."
Built With
- genai
