When it matters most, honesty is the safest answer.

ChatGPT generates text that sounds plausible — including resources that may not exist. ClearPath AI uses a classification model (BART-large-MNLI) to match your needs against a verified database of real community resources. We don't generate answers; we classify them. Every resource we recommend actually exists and has been verified.

No. ClearPath AI is designed with a zero-retention architecture. Your queries are processed in real-time and never stored on our servers. We don't track, log, or profile users. When you close the session, your data is gone — permanently.

Before any AI model runs, we check your input against a hardcoded list of crisis keywords and patterns. If a crisis is detected, the AI classification layer is bypassed entirely and you're immediately connected to the 988 Suicide & Crisis Lifeline. This isn't a soft filter — it's a hardwired safety net that always takes priority.

Every result from ClearPath AI includes an honest confidence score that tells you how certain the system is about its classification. Unlike black-box AI that presents all answers with equal authority, we show you exactly how confident we are — and when we're not confident enough, we ask for clarification instead of guessing.

Yes. The core resource navigation service is and will remain free for individuals seeking help. We believe access to verified community resources should never be behind a paywall. Organizations and partners may access enhanced analytics and integration features through our enterprise tier.

The confidence scores shown are the raw softmax probabilities returned by BART-large-MNLI on HuggingFace Inference API — unmodified, uncalibrated, exactly what the model produced. We do not inflate them, smooth them, or hide them. We have not yet built a held-out evaluation dataset to measure classification accuracy rigorously — that is on our roadmap. The crisis detection layer uses deterministic regex patterns, so its accuracy depends on the patterns we wrote (175 patterns, manually reviewed). What we can promise: the number you see is the number the model returned, not a marketing-adjusted version of it.

BART-large-MNLI classifies across 8 resource categories: Housing Assistance, Food Assistance, Mental Health, Employment, Legal Aid, Healthcare, Senior Services, and Veteran Services. Crisis Support is NOT a BART category — it is handled by a separate hardcoded regex layer that runs BEFORE BART is invoked, because crisis routing cannot depend on a probabilistic model. Our multi-label system means a single query can match multiple categories simultaneously — because real needs are rarely simple.

Our current demo supports English. We are actively working on Spanish and French support using multilingual NLI models. Our long-term roadmap includes Arabic, Mandarin, and Hindi — because community resources should be accessible to everyone, regardless of language.

When confidence drops below 70%, ClearPath AI asks a clarifying question instead of guessing. Below 50%, we escalate to a human 211 navigator. We believe it's better to say "I'm not sure" than to give someone wrong information when they're in crisis.

Every resource card includes a "Last verified" date — currently May 2026 for all entries, since we hand-curated the database ourselves during the hackathon build. We do not yet have an automated pipeline to keep this fresh. If a resource is older than 30 days (which will happen as time passes), we display a "Call to confirm" notice. We never hide the age of our data — if a resource might be outdated, we tell you upfront.

Your text input is sent over HTTPS to our classify API route, which forwards it to HuggingFace Inference API (facebook/bart-large-mnli) for zero-shot classification. We do not log input text on our servers. We do not store queries in a database. There is no user account system, no session persistence, no cookies beyond what the browser needs to render the page. HuggingFace's own data retention policy applies to their API tier — we recommend reviewing it if you have concerns. We never share, sell, or analyze input data. If you are in crisis and uncomfortable typing, call 988 directly — you do not need to use this tool.

There is no API, embed widget, or enterprise tier. ClearPath AI is a hackathon build (June 2026). The codebase is open source on GitHub — organizations are welcome to fork it, inspect it, or self-host it. We do not offer paid plans, custom configurations, or partnership agreements. If you are a nonprofit that wants to use this internally, the README has deployment instructions.