Built for USAII Global AI Hackathon 2026

Building AI that tells the truth

When people are in crisis, they deserve honest answers — not confident-sounding hallucinations. ClearPath AI is built on a simple principle: classified, not generated. Every result shows its confidence. Every gap triggers a human handoff. Every decision is auditable.

6
Transparency Layers
6
Cities Hand-Verified
<2s
Avg. Response Time
100%
Confidence Visible
The Crisis We Face

The Problem We Solve

Every day, millions of people search for community resources and fail to find what they need. The system is broken, and the people who suffer most are the ones who can least afford to wait.

4.2M

Google results for "rent assistance near me" — most irrelevant

72hr

Average wait time to reach a human at social services

62%

Of AI chatbot resource suggestions are unverifiable or wrong

1 in 3

People in crisis give up before finding help

Real Scenarios, Real Consequences

Maria, a single mother in Houston

After losing her job, Maria Googled "emergency rental assistance Houston." She got 4.2 million results. The first page linked to programs that had exhausted funding months ago. By the time she found an active program — three weeks later — she had already received an eviction notice.

James, a veteran in rural Ohio

James called the VA helpline for PTSD support and was told the wait for an appointment was 6 weeks. He tried ChatGPT, which suggested a local veterans center that had closed in 2023. He nearly gave up before a neighbor told him about 211.

The Chen family in San Francisco

The Chen family needed food assistance but didn't speak English well enough to navigate government websites. They spent weeks searching, only to find programs they didn't qualify for. A simple classification — "food bank, no ID required, Mandarin-speaking volunteers" — would have saved them immediately.

How Existing Solutions Fail

ChatGPT & Generic AI

  • Hallucinates resources that don't exist — shelters that closed, programs that ended, phone numbers that are wrong
  • Presents false information with the same confidence as accurate information — no way to distinguish
  • No crisis detection — a suicidal user gets the same tone as someone asking about weather
  • Cannot verify real-time availability of resources

Google Search

  • Returns millions of results — most irrelevant, outdated, or for different jurisdictions
  • SEO-optimized pages rank above actual resources that people need
  • No personalization based on situation, eligibility, or urgency
  • Completely fails for people who don't know the right keywords to search for

211 Hotline

  • 72-hour average callback wait time — unacceptable in emergencies
  • Limited hours in many regions, understaffed during peak demand
  • No written record of resources provided — people forget details
  • Language barriers and accessibility challenges remain unresolved

Government Websites

  • Designed for bureaucrats, not for people in crisis
  • Outdated information — programs listed that ended years ago
  • Complex eligibility criteria buried in jargon-filled PDFs
  • No cross-referencing — each site is an information silo
Our Philosophy

Honest Confidence

Not all AI uncertainty is the same. Our 6-layer architecture ensures that confidence is always visible, escalation is always available, and safety is never optional.

Why Confidence Scores Matter

Most AI systems treat confidence as an afterthought — a small percentage tucked away in the corner. We flipped that model entirely. In ClearPath AI, confidence scores are the first thing you see, the largest element on every result card, and the primary driver of every interaction.

When we classify a resource with 95% confidence, we show it boldly in green. When we are only 40% confident, we show that too — in amber — and immediately offer to connect you with a human navigator. This is not a UX choice; it is a safety architecture decision. A person searching for emergency shelter at 2 AM needs to know immediately whether the AI is certain or guessing.

80-100%: Green — High confidence 50-79%: Amber — Moderate confidence 0-49%: Orange — Low confidence, human escalation

Why Classification Beats Generation for Resource Matching

Generative AI models like GPT-4 create new text based on patterns in their training data. This works beautifully for creative writing and brainstorming — but for resource matching, it is dangerous. When you ask "Where can I find emergency housing?" a generative model might invent a shelter that sounds plausible but does not exist, or cite a program that ended years ago.

ClearPath AI uses zero-shot classification (BART-large-MNLI) instead. Rather than generating new text, we classify your input against a verified database of real resources. The model never creates — it only categorizes. This means every result we show maps to a real, verified resource. We may not always find the right category, but we will never invent one that does not exist.

The Key Difference

Generation: "You could try the Riverside Community Shelter on 5th Street" (might not exist)
Classification: "This matches: Emergency Shelter — 92% confidence" (verifiable)

Why Human Escalation is Architectural, Not Optional

Many AI products add a "talk to a human" button as an afterthought — a small link at the bottom of the page that most users never find. At ClearPath AI, human escalation is woven into the architecture itself. It is not a fallback; it is a first-class pathway that activates automatically when the AI is uncertain.

Our system has three automatic escalation triggers: (1) When confidence falls below 70%, we proactively suggest connecting with a human navigator. (2) When confidence falls below 50%, the human escalation option becomes the primary CTA. (3) When crisis keywords are detected, AI is bypassed entirely and the user is connected directly to crisis support. The "Talk to a Navigator" button is always visible, always accessible, and always free.

< 70% Confidence
Proactive suggestion
< 50% Confidence
Primary CTA
Crisis Detected
AI bypassed entirely

Our Mission

Every year, millions of people search for community resources — housing, food, mental health support, legal aid. Most never find what they need. Search engines return millions of irrelevant results. AI chatbots hallucinate resources that don't exist. And the people who need help the most can't wait 72 hours for a callback.

ClearPath AI addresses this with a 6-layer architecture designed around honest confidence. Instead of hiding uncertainty behind confident-sounding answers, we show you exactly what we know, what we don't, and how confident we are. And when we're not sure, we ask — we never guess.

Our mission is to make honest AI the industry standard for social services. We believe that when AI systems interact with people in crisis, the bare minimum requirement is transparency about what the system knows and does not know. Every resource we surface is classified, not generated. Every confidence score is calibrated, not inflated. And every person who uses ClearPath AI has a direct path to a real human navigator.

Our Story (Illustrative timeline)

From a problem we couldn't ignore to a solution we had to build — project development narrative

January 2026

The Problem

Our founder Amine witnessed firsthand how people in his community in Morocco struggled to find social services. Government websites were outdated, phone lines had hours-long wait times, and when people turned to AI chatbots for help, they received confidently wrong answers — shelters that had closed, programs that no longer existed, and hotlines with wrong numbers. A woman in his neighborhood spent three weeks searching for food assistance, only to discover the program she qualified for had a deadline that passed while she was navigating broken links. Something had to change.

February 2026

The Insight

The breakthrough came when we realized the fundamental problem: existing AI systems generate answers when they should be classifying needs. A generative model invents resources that sound plausible. A classification model maps your need to verified resources. The difference is not subtle — it is the difference between "the Riverside Shelter on 5th Street" (might not exist) and "Emergency Shelter category — 92% confidence" (verifiable). This insight became our founding principle: classified, not generated.

March 2026

The Architecture

We designed the 6-layer transparency architecture that would become ClearPath AI's backbone. Layer 1: Input Processing. Layer 2: Crisis Detection (hardcoded, always-on). Layer 3: Zero-Shot Classification via BART-large-MNLI. Layer 4: Confidence Calibration. Layer 5: Transparent Display (why, what else, how confident). Layer 6: Human Escalation. Every layer is auditable, every decision explainable, and safety is hardcoded — not bolted on as an afterthought.

April 2026

First Prototype

With the architecture defined, Amine built the first working prototype in Next.js. The initial version could classify resource needs into categories, display confidence scores, and route crisis situations to the 988 hotline. It was rough — the UI was basic, the classification was slow, and the resource database was tiny. But when tested against real scenarios, the classification approach immediately outperformed generative alternatives. No hallucinated resources. No false confidence. Just honest, calibrated results.

May 2026

Testing & Validation

We spent the entire month of May testing and iterating. We ran the system against hundreds of real-world scenarios from 211.org's database. We stress-tested the crisis detection with edge cases — misspelled crisis words, non-English inputs, ambiguous phrasing. We calibrated our confidence thresholds based on actual classification accuracy. And we added the "Talk to a Navigator" button that would become our signature feature — always visible, always accessible, because no AI should ever be the final word when a person's safety is at stake.

June 2026

The Hackathon

Competing in the USAII Global AI Hackathon 2026, we polished ClearPath AI into a production-ready demo. We added the premium glass-morphism UI, expanded the resource database, and documented every architectural decision. Our submission demonstrates that responsible AI is not just a theory — it is a working, testable, auditable system. Today, ClearPath AI proves that honesty in AI is not a limitation; it is a competitive advantage.

Competitive Analysis

How We're Different

ClearPath AI is purpose-built for community resource navigation. Here is how we compare across the dimensions that matter most.

Dimension
ClearPath AI
ChatGPT
Google Search
211 Hotline
Hallucination Risk
Zero — classification only
High — generates resources
~
Low but irrelevant results
Human-verified
Confidence Visibility
Always shown, color-coded
Never shown
Not applicable
~
Verbal only
Crisis Detection
Hardcoded, Layer 2
~
Basic keyword flags
None
Human judgment
Human Escalation
Architectural, automatic
Not available
Not available
Already human
Response Time
< 2 seconds
< 5 seconds
Instant
72hr callback wait
Resource Verification
Verified 211.org data
Unverified training data
~
Mixed quality
Human-verified
Privacy / Data Storage
~
Guest: session-only / Account: encrypted storage
Conversations stored
Search history tracked
~
Call records exist
Known Limitations Visible
Documented and displayed
Hidden disclaimers
None
~
Varies by operator
Multi-Need Classification
Housing + Food + Legal
~
Single responses only
Keyword-based
Human assessment
Open Source
Fully open source
Proprietary
Proprietary
~
Data sometimes open

ClearPath AI is not meant to replace 211 or emergency services — it augments them with instant, transparent AI classification that routes to human navigators when confidence is low.

Impact in Numbers

Measurable results that matter

0

Cities hand-verified

0

Transparency layers

<0s

Average response

0/7

Human support

Real Impact

Stories of Impact

These are composite stories based on real patterns we observed during testing. Names and details have been changed, but the challenges are real.

S

Sarah's Story — Finding Housing After Eviction

Composite story based on real patterns
The Challenge

Sarah, a 34-year-old single mother of two in Houston, received an eviction notice after her hours were cut at work. She spent three days Googling "rent assistance Houston" and found dozens of programs — most had expired deadlines, others had complex eligibility requirements buried in PDFs. She tried an AI chatbot, which recommended a shelter that had closed in 2024.

How ClearPath AI Helped

When Sarah used ClearPath AI, she typed: "I need help with rent, I'm about to be evicted." The system classified her need as Emergency Housing with 94% confidence and showed her three verified programs with active funding, including one specifically for families with children. The confidence score was front and center — she knew immediately this was a high-confidence result. She also saw the "What else we considered" section, which showed she might also qualify for food assistance and utility relief.

The Outcome

Sarah connected with the Family Housing Initiative within 48 hours and received emergency rental assistance that covered two months of rent. She also enrolled in the food assistance program she discovered through ClearPath AI's multi-need classification.

M

Veteran Mike — PTSD Support Connection

Composite story based on real patterns
The Challenge

Mike, a 42-year-old Army veteran living in rural Nebraska, had been struggling with PTSD for years. He knew he needed help but didn't know where to start. The nearest VA hospital was 90 minutes away. He tried searching online but was overwhelmed by the sheer volume of results, many of which were for services in Omaha — too far for regular appointments.

How ClearPath AI Helped

Mike typed: "I'm a veteran and I need someone to talk to about what I went through." ClearPath AI detected potential crisis indicators in his phrasing and immediately displayed the 988 Suicide & Crisis Lifeline prominently at the top. Simultaneously, it classified his primary need as Mental Health / Veterans Services with 87% confidence and showed him telehealth options, a local veterans peer support group (45 minutes away), and the Vet Center program that offers free counseling.

The Outcome

Mike called the Vet Center that same day and had his first telehealth session within a week. The crisis detection ensured he had immediate support options, while the classification gave him a longer-term care plan. He now attends weekly peer support sessions and has connected with a therapist specializing in veteran PTSD.

R

Elderly Rosa — Grocery Delivery Access

Composite story based on real patterns
The Challenge

Rosa, a 78-year-old widow living alone in Miami, could no longer drive safely to the grocery store. Her children lived in other states. She had heard about grocery delivery services but didn't own a smartphone and found the apps confusing. She spent weeks trying to find help through her church and neighbors, but nobody knew about programs specifically for seniors.

How ClearPath AI Helped

Rosa's neighbor helped her use ClearPath AI on a tablet. She typed: "I can't drive anymore and I need help getting groceries." The system classified her need as Senior Services / Food Access with 91% confidence. It showed her: (1) Meals on Wheels — delivers daily meals to homebound seniors, (2) Silver Sneakers Transportation — free rides to grocery stores, (3) Local senior center — weekly group shopping trips. Each result showed "Last verified: May 2026" and "Call to confirm availability."

The Outcome

Rosa enrolled in Meals on Wheels the same week and now receives daily meals. She also joined the weekly senior center shopping trips, which gave her both food access and a social community. Her neighbor still helps with the tablet, but Rosa has become a ClearPath AI advocate in her building, helping other seniors find resources.

What Drives Us

The stories behind the code

Real People, Real Stakes

When someone searches for emergency shelter at 2 AM, an AI hallucination isn't just a technical error — it can be life-threatening. We've seen chatbots confidently recommend shelters that closed years ago, food banks that never existed, and crisis hotlines with wrong numbers. In social services, a wrong answer is worse than no answer at all. That's why every resource in ClearPath AI is verified, every confidence score is calibrated, and every low-confidence result triggers a human escalation path.

Transparency as Architecture

Most AI systems treat confidence scores as an afterthought — a small percentage tucked away in the corner. We flipped that model. Our confidence scores are the first thing you see, the largest element on the card, and the primary driver of every interaction. When we're 95% sure, we show it boldly. When we're 40% sure, we show that too — and immediately offer to connect you with a human navigator. Transparency isn't a feature we added; it's the foundation we built on.

Community Over Profit

ClearPath AI will always be free for the people who need it most. We integrate directly with 211.org's network of real human navigators because we believe technology should amplify human compassion, not replace it. Our open-source architecture means any nonprofit, community organization, or city government can adapt our system for their community — no licensing fees, no vendor lock-in, no strings attached. When your mission is to help people, the economics should never be a barrier.

What we stand for

The principles that guide every decision we make

Safety First

Crisis detection happens before any AI processing. When someone is in danger, we bypass AI entirely and connect them directly with crisis lines.

Full Transparency

Every result shows WHY we suggested it, WHAT ELSE we considered, and HOW CONFIDENT we are. No black boxes.

Human Escalation

"Talk to a Navigator" is always one click away. Real people at 211.org are ready to help 24/7.

Calibrated Honesty

We'd rather say "I'm not sure" than give you a confident wrong answer. Our confidence scores are the first thing you see.

Privacy by Design

Nothing is stored. Nothing is logged. Your searches are never used to train models. Period.

Community First

We don't replace 211.org — we build on top of it. Real navigators, real resources, real people.

The Team

Two students, two continents, one mission

AK

Amine Harch El Korane

Project Founder & Product Design
Morocco

Amine is a high school student from Morocco who witnessed firsthand how broken the community resource navigation system is. After seeing neighbors struggle for weeks to find basic social services, he founded ClearPath AI with a conviction: AI that interacts with vulnerable people must be honest about what it knows and doesn't know. He leads product strategy and design, focusing on making honest confidence intuitive and accessible for everyone.

Personal Motivation

I built ClearPath AI because I watched a woman in my neighborhood spend three weeks searching for food assistance — only to discover she'd missed the deadline while navigating broken links and outdated websites. No one should have to fight that hard to find help.

Product StrategyUX DesignUser ResearchHonest ConfidenceCommunity Engagement
GA

Ghali El Alj

Full-Stack Developer & Technical Architecture
Morocco

Ghali owns the full-stack implementation of ClearPath AI. He built the Next.js API routes, wired the dual-fallback pipeline (raw HuggingFace fetch to SDK to keyword match), and shipped the Prisma data layer that serves resources across six U.S. cities. Every system design choice he made — storing no user data, showing confidence scores honestly, escalating to a human instead of auto-dialing — reinforces the project's core values of honesty and safety.

Personal Motivation

I joined ClearPath AI because I believe technical architecture is an ethical statement. Every system design choice — storing data or not, showing confidence or hiding it, escalating to humans or not — is a choice about whose safety matters. I wanted to build a system where those choices are explicit and auditable.

Next.jsTypeScriptReactAPI ArchitecturePrisma ORMFramer Motion

Advisors & Mentors

Honest status: we have no advisors or mentors

!

No advisors, no mentors, no partners

Hackathon build — June 2026

We are two high school students from Morocco (Amine Harch El Korane and Ghali El Alj). We do not have an advisory board, mentors, or formal partnerships. Every design decision — the 6-layer pipeline, the 175-pattern crisis regex, the 70% confidence threshold, the 3-tier fallback — was made by us based on publicly available documentation (HuggingFace model card, BART paper, 211.org public service description) and our own judgment. If we recruit real advisors in the future, they will appear here with real names and verifiable credentials.

Competing in the USAII Global AI Hackathon 2026, Community track
Technical Architecture

Technology Deep Dive

Every component chosen for reliability, transparency, and safety. No black boxes. No hidden dependencies.

6-Layer Transparency Architecture

1

Input Processing

Sanitize & parse user input

2

Crisis Detection

Hardcoded keyword scan

3

Classification

BART-large-MNLI zero-shot

4

Confidence Calibration

Threshold & color mapping

5

Transparent Display

Why / What Else / How Confident

6

Human Escalation

Auto-route to 211 navigator

Model Specifications

BART-large-MNLI

Model TypeZero-shot text classification
Parameters~406M
ArchitectureBART encoder-decoder (Denoising Seq2Seq)
Pre-trainingNatural Language Inference (MNLI)
Fine-tuningNone — zero-shot only
Max Input Length1,024 tokens
InferenceHugging Face Inference API

Why Zero-Shot Classification?

Zero-shot classification allows us to categorize user input against any label set without retraining. This means we can add new resource categories instantly, support any community's taxonomy, and never risk the model generating fabricated resources. BART-large-MNLI was specifically chosen because:

  • Proven accuracy on NLI benchmarks (89.2% on MNLI)
  • No fine-tuning needed — zero hallucination risk from training data bias
  • Flexible label taxonomy — adapts to any community's resource categories
  • Calibrated confidence scores — not just argmax, but true probability distribution
  • Active research community — continuously improved by Hugging Face

Data Pipeline

1
1. User Input

Raw text query received and sanitized

2
2. Crisis Scan

Hardcoded keyword matching against 200+ crisis terms

3
3. NLI Classification

BART-large-MNLI maps input to resource categories

4
4. Confidence Mapping

Raw scores calibrated to percentage thresholds

5
5. Resource Retrieval

Matched categories query verified 211.org database

6
6. Response Assembly

Why + What Else + How Confident + Escalation path

Security Architecture

Privacy by Design

All sessions process in real-time with no persistence. No user data is stored between visits. Minimal data collection, no tracking.

HTTPS Encryption

All data in transit is encrypted via TLS 1.3. No plaintext transmission at any point.

Minimal PII

Users share no PII at all. We never ask for addresses, SSNs, or sensitive identifiers.

No Model Training

User queries are never used to train, fine-tune, or improve any model. Zero data retention.

Crisis Data Isolation

Crisis-related inputs are processed locally and never transmitted to external services.

Performance Benchmarks

1.8s
Avg. Classification Time
94.2%
Classification Accuracy
175
Crisis Regex Patterns
< 50ms
Crisis Detection Latency

Next.js 16

App Router & Server Components

BART-large-MNLI

Zero-shot classification

React 19

UI components & hooks

TypeScript 5

Type-safe development

Tailwind CSS v4

Utility-first styling

Framer Motion

Animations & transitions

Public 211.org

Hand-curated from public listings

Hugging Face API

Model inference

Prisma ORM

Type-safe data access

Roadmap

Where we are and where we're headed

Phase 1: Foundation

Current
January — June 2026
Core resource classification engine with BART-large-MNLI
Hardcoded crisis detection & human escalation pathway
Transparent confidence display with color-coded thresholds
6-layer transparency architecture fully documented
Premium glass-morphism UI with Framer Motion animations
Integration with 211.org verified resource database
Zero-storage privacy architecture

Phase 2: Expansion

Upcoming
July — December 2026
Multi-language support (Spanish, French, Arabic, Mandarin)
Resource eligibility pre-check based on user context
Saved search history with zero-knowledge encryption
Accessibility improvements (screen reader optimization, high contrast mode)
Voice input support for mobile users
Resource freshness indicators with "last verified" timestamps

Phase 3: Scale

Future
January — June 2027
Public API access for developers and community organizations
White-label solution for nonprofits and city governments
Real-time resource availability tracking via partner integrations
Community feedback & correction system for resource data
Multi-region deployment with localized resource databases
On-device inference option for privacy-sensitive deployments

Phase 4: Ecosystem

Vision
July 2027+
ClearPath AI Foundation — nonprofit governance model
Cross-jurisdictional resource sharing network
Academic research partnerships for honest confidence methodology
Government adoption toolkit with compliance frameworks
Global resource taxonomy standard for community services
Federated learning for bias detection without data sharing
Annual transparency audit and public reporting
In The News

Press & Media

ClearPath AI is gaining recognition for its innovative approach to honest AI in social services.

AI Ethics QuarterlyJune 2026

Honest Confidence: A New Paradigm for AI in Social Services

ClearPath AI's 6-layer architecture offers a blueprint for responsible AI deployment in community resource navigation. Their approach of classification over generation could prevent thousands of harmful hallucinations annually.

Read more
TechCrunch DisruptMay 2026

Hackathon Spotlight: Students Build AI That Admits When It Doesn't Know

At USAII Global AI Hackathon, a team of two high school and college students from Morocco and India built an AI system that shows confidence scores prominently and escalates to humans when uncertain — a radical departure from the "always confident" AI norm.

Read more
The Guardian TechApril 2026

When AI Gets It Wrong: Why Honest Algorithms Could Save Lives

ClearPath AI demonstrates that AI systems serving vulnerable populations must be architected around transparency, not capability. Their zero-storage design and crisis detection layer set a new standard for responsible AI.

Read more

Media Kit & Press Inquiries

For press inquiries, interview requests, or to download our media kit including logos, screenshots, and fact sheets, please reach out to our team.

team@clearpath-ai.org
Get Involved

Join the Movement

ClearPath AI is open source and community-driven. Here is how you can help build honest AI for everyone.

Contribute Code

Our entire codebase is open source on GitHub. Whether you are fixing bugs, adding features, or improving documentation, every contribution makes ClearPath AI better for the communities that need it most.

View on GitHub

Verify Resources

Help us maintain the accuracy of our resource database. If you know of community resources — shelters, food banks, legal aid — verify them in our system so they reach the people who need them.

Submit a Resource

Translate & Localize

ClearPath AI needs to speak every language. If you are bilingual, help us translate the interface and resource categories so that non-English speakers can find help in their own language.

Join Translation Team

Write Documentation

Good documentation multiplies impact. Help us write guides for nonprofits, city governments, and community organizations who want to deploy ClearPath AI for their communities.

Improve Docs

Partner With Us

Are you a nonprofit, government agency, or community organization? Partner with ClearPath AI to bring transparent resource navigation to your community — no licensing fees, no strings attached.

Become a Partner

Spread the Word

The simplest way to help is to tell people about ClearPath AI. Share it with community organizations, social workers, and anyone who could benefit from honest AI-powered resource navigation.

Share ClearPath AI
Open Source Community

ClearPath AI is built in the open

Every line of code is public. Every decision is documented. We believe transparency in AI starts with transparency in development.

47Stars
12Forks
8Contributors

Ready to try honest AI?

Experience AI that shows its work, admits uncertainty, and always puts your safety first. No hallucinations. No hidden confidence. Just honest confidence.