| # Borderless Field Notes | |
| ## Why This Exists | |
| Immigration research is high-stakes, fragmented, and exhausting. A person trying | |
| to move for work, study, family, or safety often has to compare government pages, | |
| embassy notes, point calculators, processing-time pages, blogs, and anecdotes. | |
| Borderless is built for the moment before someone pays a lawyer: "What countries | |
| are realistically worth researching for my profile?" | |
| ## Why A Small Model Fits | |
| The hard part is not memorizing every immigration rule. Policies change too | |
| often. A small model works well when it is used as a planner and synthesizer: | |
| - It turns an everyday-language profile into a concrete research plan. | |
| - It calls search and scraping tools for current official sources. | |
| - It compares tradeoffs in plain language. | |
| - It drives the globe so geography is part of the reasoning, not decoration. | |
| This is a better fit than a larger model guessing from stale training data. | |
| ## What Changed After Prototype Testing | |
| The first prototype was a pure chat with visible tool JSON and an editable system | |
| prompt. It worked, but a new user had to know what to ask and judges had to read | |
| through debugging details. | |
| The award-ready version adds: | |
| - A guided intake form that creates a complete research prompt. | |
| - Demo personas for quick evaluation. | |
| - A stricter research protocol focused on official sources. | |
| - Source-quality labels in search results. | |
| - Country-profile metadata for shortlist and globe grounding. | |
| - Friendlier tool traces that explain progress without raw JSON noise. | |
| - Globe loading and empty states so the visual panel is useful immediately. | |
| - Local JSONL trace logging that can be shared as hackathon evidence. | |
| ## Responsible Use | |
| Borderless is research support, not legal advice. The agent is instructed to cite | |
| official sources for eligibility, documents, fees, timelines, and application | |
| steps. When official data is missing or a tool fails, the answer should say what | |
| needs verification instead of inventing details. | |
| ## Hackathon Award Fit | |
| - Backyard AI: a practical tool for a real, specific life decision. | |
| - Best Agent: multi-step tool use, source quality, partial synthesis, and globe | |
| updates are all model-driven. | |
| - Best Demo: the guided form plus globe shortlist creates an immediate visual | |
| story. | |
| - Sharing is Caring: JSONL traces can be published with the Space or linked in a | |
| writeup. | |
| - Field Notes: this report documents the product choices and constraints. | |