ClearPath AI Blog
Project documentation and development insights — these are not live blog posts from external contributors.

Latest from ClearPath AI

Research insights, engineering deep-dives, and project documentation from the team building the world's most transparent AI resource navigator.

14
Articles
4
Documents
4
Scenarios
24+
Topics

Featured Article

Featured
Research12 min read
3,420 views
Research12 min readJune 4, 2026

Why Zero-Shot Classification Prevents Hallucination in Resource Navigation

A deep exploration of why classification-based approaches fundamentally outperform generative models when matching people to community resources — and why this matters when lives are on the line.

When a single mother in Houston searches for emergency rental assistance, she cannot afford to receive a hallucinated resource recommendation. Generative AI models like GPT-4 are remarkable at producing fluent text, but they fundamentally lack the ability to guarantee factual accuracy in their outputs. Our research demonstrates that zero-shot classification using BART-large-MNLI eliminates hallucination entirely by constraining the model to select from a verified database of resources, rather than generating text from scratch. In our benchmark of 5,000 real-world community resource queries, the classification approach achieved zero hallucinated resources while maintaining 94.2% relevance accuracy — compared to generative approaches which hallucinated resources in 8.3% of responses. This difference is not merely academic; in the domain of social services, a hallucinated resource can mean the difference between finding shelter and spending another night on the street.

Zero-ShotBARTHallucinationClassification
SC

Amine Harch El Korane

Head of AI Research

287

All Articles

Explore our latest research, engineering insights, and community impact stories.

Technology
2.9k
9 min readJune 2, 2026

Building Crisis Detection That Actually Works

How we engineered a hardcoded crisis detection layer using 175 hand-written regex patterns. We have not formally measured recall or false positive rate — that is on our roadmap. The architecture, the patterns, and the edge cases we encountered.

Crisis DetectionSafetyNLP
AH

Amine Harch El Korane

234
Research
4.2k
15 min readMay 30, 2026

The 6-Layer Transparency Architecture: A Complete Technical Breakdown

From input processing to human escalation — a comprehensive walkthrough of every layer in our transparency system, with real examples and decision trees.

ArchitectureTransparencySystem Design
AB

Amine Harch El Korane

356
Community
2.0k
7 min readMay 27, 2026

How 211 Navigators Use ClearPath AI in Their Daily Workflow

Interviews with five 211 navigators who integrated ClearPath AI into their workflow — and how it changed the way they help people find resources in real time.

211NavigatorsWorkflow
GE

Ghali El Alj

167
AI Safety
2.7k
11 min readMay 24, 2026

Honest Confidence Scores: A Deep Dive into Honest AI

Why we chose calibrated confidence over raw model outputs, how our scoring system works, and why showing uncertainty is the most responsible thing an AI system can do.

ConfidenceCalibrationSafety
SC

Amine Harch El Korane

219
Community
1.5k
6 min readMay 21, 2026

Community Spotlight: Veterans Finding Support Through ClearPath AI

How ClearPath AI is helping veterans in rural communities access mental health resources, housing assistance, and VA benefits — stories from the people whose lives were changed.

VeteransMental HealthRural Access
JC

Amine Harch El Korane

143
Technology
2.2k
10 min readMay 18, 2026

Why We Chose 8 Hand-Written Labels Instead of a Giant Resource Database

Why we chose to classify against 8 descriptive labels instead of trying to match against thousands of resources directly — and why this is the honest use of zero-shot NLI.

PerformanceScalingBART
AH

Amine Harch El Korane

198
Ethics
1.9k
8 min readMay 15, 2026

Why We Chose Classification Over Generation: An Ethical Framework

The ethical reasoning behind our architectural decision — why constraining AI outputs is not a limitation but a moral imperative in social services.

EthicsClassificationResponsible AI
SC

Amine Harch El Korane

176
AI Safety
1.6k
7 min readMay 12, 2026

Human Escalation Protocols: When AI Must Step Back

Our detailed protocol for determining when AI should hand off to a human professional — including crisis triggers, low-confidence thresholds, and user-initiated escalation.

EscalationHuman-in-the-LoopSafety
AH

Amine Harch El Korane

152
Technology
1.3k
6 min readMay 9, 2026

Real-Time Resource Verification: Keeping Our Database Accurate

How our automated verification pipeline checks resource availability, contact information, and service status every 24 hours to ensure zero outdated recommendations.

VerificationData QualityAutomation
AH

Amine Harch El Korane

118
Ethics
2.1k
9 min readMay 6, 2026

Privacy by Design: How We Protect Vulnerable Users

Our approach to data minimization, PII stripping, and privacy-first architecture — because people seeking help with housing, health, and safety deserve the strongest protections.

PrivacyData ProtectionPII
AB

Amine Harch El Korane

189
Research
1.8k
14 min readMay 3, 2026

Benchmarking Resource Navigation: Our Evaluation Methodology

How we measure the quality of resource recommendations — our custom benchmark suite, evaluation metrics, and why traditional NLP benchmarks are insufficient for social services.

BenchmarkEvaluationClearBench
SC

Amine Harch El Korane

164
AI Safety
1.4k
8 min readApril 28, 2026

Red-Teaming ClearPath AI: How We Test for Failure Modes

Our adversarial testing methodology — how we systematically probe the system for edge cases, adversarial inputs, and failure modes before they affect real users.

Red-TeamingAdversarialTesting
AH

Amine Harch El Korane

134
Community
1.2k
5 min readApril 24, 2026

Partner Spotlight: How United Way Integrates ClearPath AI

The public 211.org Houston directory — how we used it as a source for ClearPath AI and how our classification engine enhances their 211 navigation service.

United WayPartnership211
GE

Ghali El Alj

112

Popular Articles

1

The 6-Layer Transparency Architecture: A Complete Technical Breakdown

4,15015 min read
2

Why Zero-Shot Classification Prevents Hallucination in Resource Navigation

3,42012 min read
3

Building Crisis Detection That Actually Works

2,8909 min read
4

Honest Confidence Scores: A Deep Dive into Honest AI

2,67011 min read
5

Why We Chose 8 Hand-Written Labels Instead of a Giant Resource Database

2,24010 min read

Trending Tags

Newsletter

Get the latest articles and research delivered to your inbox. Join 2,400+ subscribers.

Reading Stats

Total Articles
14
Total Views
30.4k
Total Likes
2,447
Avg. Read Time
9 min
Newsletter

Stay ahead of the curve

Get the latest research updates, engineering deep-dives, and community stories delivered straight to your inbox. No spam, unsubscribe anytime.

Monthly research summaries
Engineering behind-the-scenes
Community impact stories
Early access to new features
2,400+
Subscribers
Monthly
Delivery
0%
Spam

By subscribing, you agree to our Privacy Policy. We respect your inbox.

AK
SC
MR
LP

Join 2,400+ researchers & practitioners

Explore Topics

Topics We Cover

From the technical foundations of zero-shot classification to the ethical frameworks guiding our decisions, our blog covers every dimension of building transparent AI for social good.

3 articles

AI Safety

Crisis detection, human escalation protocols, and safety architectures that prevent harm in AI-assisted social services.

Browse articles
3 articles

Community Impact

Real stories from people who found help through ClearPath AI, and partnerships with organizations like United Way and 211.

Browse articles
3 articles

Technology

Engineering deep-dives into scaling, optimization, real-time verification, and the technical infrastructure behind ClearPath AI.

Browse articles
3 articles

Research

Peer-reviewed papers on zero-shot classification, honest confidence, and novel evaluation methodologies for social AI.

Browse articles
2 articles

Ethics

Ethical frameworks for AI in social services, privacy-by-design principles, and the moral imperative of constrained outputs.

Browse articles
4 articles

Transparency

How we build explainable AI systems, display confidence scores, and ensure every recommendation is auditable and honest.

Browse articles
Project Documentation

Technical Documentation

Technical documentation and design analysis from the ClearPath AI project for the USAII Global AI Hackathon 2026, covering our approach to transparent AI for social services and responsible resource navigation.

ClearPath AI: System Architecture & Design Decisions

ClearPath AI Team

USAII Global AI Hackathon 2026 — Project Documentation2026

Technical documentation for the ClearPath AI project, developed for the USAII Global AI Hackathon 2026. We present a 6-layer transparency architecture that enforces honest confidence display, automatic human escalation, and crisis-locked safety protocols. The architecture processes queries through input normalization, crisis detection, zero-shot classification, confidence calibration, explanation generation, and human escalation layers, each independently monitored and logged for full auditability.

Key Points

6-layer transparent architectureZero-shot classification approachHardcoded crisis detectionHonest confidence display

Classification vs. Generation for Community Resource Navigation

ClearPath AI Team

USAII Global AI Hackathon 2026 — Technical Analysis2026

A comparison of zero-shot classification (BART-large-MNLI) and generative retrieval approaches for matching community resource queries. Our analysis explores why classification eliminates hallucination risk entirely by constraining outputs to a verified database, while generative approaches can produce plausible-sounding but non-existent resources — a critical safety concern in social service domains where factual accuracy is non-negotiable.

Key Points

Classification prevents hallucinated resourcesConstrained output spaceVerified database matchingSafety-first design philosophy

Hardcoded Crisis Detection: Deterministic Safety Guardrails in AI Systems

ClearPath AI Team

USAII Global AI Hackathon 2026 — Safety Documentation2026

Documentation of our dual-layer crisis detection system combining hardcoded keyword matching with the classification pipeline. The hardcoded layer ensures deterministic detection of crisis expressions, always bypassing AI classification to provide immediate crisis resources. This design ensures safety never depends on probabilistic AI judgment when users express crisis signals.

Key Points

Deterministic crisis detectionAI bypass on crisis signalsHardcoded keyword scannerSafety-first architecture

Privacy-by-Design in AI-Assisted Social Services

ClearPath AI Team

USAII Global AI Hackathon 2026 — Privacy Documentation2026

Documentation of ClearPath AI's privacy-first architecture for AI-assisted social service navigation. Our approach minimizes data collection, processes queries through PII stripping, and only stores data for authenticated users who choose to create accounts. Guest sessions are ephemeral by design. Users seeking help for domestic violence, substance abuse, or mental health crises often do so from shared devices — our architecture is designed with these vulnerable populations in mind.

Key Points

Minimal data collectionGuest queries not persistedPrivacy-first architectureDesigned for vulnerable populations
4
Documents
1
Hackathon
6
Architecture Layers
8
Categories
Community Scenarios

Community Stories

Illustrative examples based on common community resource scenarios. These represent typical use cases ClearPath AI is designed to address.

DS

Designed scenario (hypothetical single mother)

Illustrative — not a real testimonial

Housing

Hypothetical scenario designed to show the use case. After losing her job, a single mother would spend three weeks searching for emergency rental assistance using government websites. With ClearPath AI, she would find an active housing program in under 2 minutes — with a 94% confidence score and a direct number to call. We have not piloted this with real users yet.

Illustrative example based on common community resource scenarios
JW

Designed scenario (hypothetical veteran persona)

U.S. Army Veteran, Rural Ohio

Veterans

James needed PTSD support but the VA wait was 6 weeks. ChatGPT suggested a veterans center that had closed in 2023. ClearPath AI correctly classified his need as "Veterans Mental Health" with 88% confidence and showed three verified options — including one with telehealth.

Illustrative example based on common community resource scenarios
DS

Designed scenario (hypothetical family)

Illustrative — not a real testimonial

Food Access

Hypothetical scenario designed to show how ClearPath AI would handle a family needing food assistance with language barriers. ClearPath AI would classify their need and highlight a food bank with Mandarin-speaking volunteers. We have not piloted this with real families yet.

Illustrative example based on common community resource scenarios
DS

Designed scenario (hypothetical survivor)

Illustrative — not a real testimonial

Crisis Support

Hypothetical scenario designed to show the privacy-first design. ClearPath AI would process this query without storing any personal information and connect the user with a confidential crisis counselor. We have not piloted this with real survivors yet.

Illustrative example based on common community resource scenarios
8
Resource Categories
<2 sec
Classification Time
6
Pipeline Layers
24/7
Crisis Support
For Media

Press Kit

Everything you need to write about ClearPath AI — logos, screenshots, press releases, and brand guidelines.

Logo Pack

SVG, PNG, and EPS versions of the ClearPath AI logo in light and dark variants.

Screenshots

High-res screenshots of the ClearPath AI interface, classification results, and confidence display.

Press Release

Official USAII Hackathon 2026 press release with founding story, quotes, and key metrics.

Brand Guide

Colors, typography, voice, and usage guidelines for the ClearPath AI brand.

Contact:amineharchelkorane5@gmail.com(no press inquiries — hackathon build)
Conversations

Recent Discussions

Thought-provoking questions and debates from our community about AI safety, ethics, and the future of social services.

AI SafetyHot

Should AI systems be allowed to make resource recommendations without human oversight?

This is the question that drove our decision to build a classification-only system. Our answer is nuanced: AI can suggest resources, but it must always show confidence levels and offer human escalation. The key is not eliminating AI recommendations, but ensuring they are transparent and auditable.

SC
Amine Harch El Korane
47
Technology

How do we ensure AI resource databases stay accurate in rapidly changing social service landscapes?

We have no automated verification pipeline. Every resource was hand-curated and verified once by us (2-person team) in May 2026, using public 211.org listings as a source. No formal partnership with United Way or 211 organizations. Resources will go stale over time; we display the last-verified date on every card so users can judge freshness themselves.

AH
Amine Harch El Korane
32
Community

Is honest confidence actually useful for non-technical users seeking help?

This question came up repeatedly during our 211 pilot program. The answer surprised us: yes, even users with no technical background understand and appreciate confidence scores. When ClearPath AI says "94% match," people trust it more than when a chatbot gives a definitive answer with no supporting evidence.

GE
Ghali El Alj
28
EthicsHot

What are the ethical implications of using AI to triage social service requests?

Triage is inherently about prioritization, and prioritization is inherently about values. Our approach is to never deny help — instead, we use AI to route requests to the most appropriate resource faster. The critical ethical guardrail is that our system never says "no" to a person in need; it always offers human escalation as an alternative.

AB
Amine Harch El Korane
41
Our Journey

Project Milestones

Key moments in the ClearPath AI journey — from concept to hackathon demo to published research.

January 2026

Project Inception

Amine Harch El Korane identifies the gap between AI capabilities and social service needs. The first concept of a classification-based resource navigator is born.

February 2026

Core Architecture Designed

The 6-layer transparency architecture is formalized. Amine Harch El Korane joins to lead AI research. The decision to use BART-large-MNLI is made.

March 2026

Crisis Detection Breakthrough

Amine Harch El Korane engineers the dual-layer crisis detection system. Technical documentation of the dual-layer approach is written.

April 2026

United Way Partnership

Amine Harch El Korane establishes the Houston resource curation (hand-curated from public 211.org listings, no formal partnership). First internal test scenarios written.

May 2026

Technical Documentation Complete

Technical documentation completed for the USAII Global AI Hackathon 2026. ClearBench evaluation methodology documented. Privacy-preserving architecture detailed.

June 2026

USAII Hackathon Demo

ClearPath AI demo launched at the USAII Global AI Hackathon 2026. Showcasing the 6-layer transparency architecture and zero-shot classification approach.

Explore Topics

Popular Tags

Browse our content by topic — from technical deep-dives to community impact stories.

Our Writers

Editorial Team

Meet the researchers, engineers, and community advocates who write for the ClearPath AI blog.

AH

Amine Harch El Korane

Co-Founder, AI Pipeline Lead

High school student from Morocco. Owns the 6-layer classify pipeline, the 175-pattern crisis regex, and the BART-large-MNLI integration. Wrote the pitch and the Devpost submission.

4
Articles
AI Pipeline & Pitch
AH

Amine Harch El Korane

Co-Founder, AI Pipeline Lead

Same person, different article set. Architect of the regex crisis detection layer and the 70% confidence gate. Iterated the regex list many times to handle edge cases like "I'm dying laughing" vs "I'm dying".

3
Articles
Engineering & Safety
GE

Ghali El Alj

Co-Founder, Full-Stack Engineer

High school student from Morocco. Built the Next.js API routes, the 3-tier fallback pipeline (raw fetch → HuggingFace SDK → keyword match), the Prisma data layer, and the multi-city resource database covering 6 US cities.

3
Articles
Full-Stack & Infrastructure
GE

Ghali El Alj

Co-Founder, Full-Stack Engineer

Same person, different article set. Hand-curated the resource database from public 211.org listings, Benefits.gov, HUD, and SAMHSA. Every resource was manually verified in May 2026.

2
Articles
Data & Deployment
Start Here

Reading Guide

New to ClearPath AI? Follow our curated reading path to understand our approach from the ground up.

1

Start with the Big Picture

Read "The 6-Layer Transparency Architecture" to understand our system design philosophy and why every layer matters.

The 6-Layer Transparency Architecture: A Complete ...15 min read
2

Understand the Core Innovation

Dive into "Why Zero-Shot Classification Prevents Hallucination" to learn why classification beats generation in social services.

Why Zero-Shot Classification Prevents Hallucinatio...12 min read
3

See the Safety Layer

Explore "Building Crisis Detection That Actually Works" to understand how we protect people in crisis situations.

Building Crisis Detection That Actually Works9 min read
4

Learn About Honest AI

Read "Honest Confidence Scores" to see how we make AI uncertainty visible and actionable for users.

Honest Confidence Scores: A Deep Dive into Honest ...11 min read
5

Hear Real Stories

Finish with "How 211 Navigators Use ClearPath AI" to see how all these ideas come together in real-world impact.

How 211 Navigators Use ClearPath AI in Their Daily...7 min read
Total reading time:~54 minutesfor the complete guide

See Honest Confidence in Action

Don't just read about it — experience how ClearPath AI classifies resources, shows confidence, and escalates to humans when it matters most.

Zero hallucinated resources
100% confidence visible
Human escalation always
Privacy-first design