Yellow Pages

Screenshots, not summaries - Deep research agent for nonprofit social service monitoring.

About Yellow Pages

For Sprint 1 of my bootstrapped incubator, I helped nonprofits keep their social service directories up to date. Currently they spend hours going through doctors' or pro bono legal websites to make sure they're directing clients to the right contact info, but it leads to typos/duplicates, preventing successful referrals (just over half in one nonprofit) and people getting the care they need.

I built Yellow Pages, a website to have LLMs monitor and summarize website changes, de-duplicate their existing data, and verify new information via screenshots with a from-scratch built computer-use deep research agent 10x faster and 3x cheaper than OpenAI's since other deep research tools were too slow, unreliable, and expensive for contact details.

Since my February Agents for Nonprofits hackathon, 2 nonprofits have been helped with the features and now a 3rd is using the actual site, saving hundreds of hours.

Watch Demo

Demo Video

Problem

Nonprofits manually re-check 5k+ websites every quarter to keep contact info accurate. % of successful referrals bottlenecked by directory lacking emails that existed on provider sites.

Solution

All-in-one website with autonomous agents that notify of semantic website changes, de-duplication algorithms, and screenshot proof instead of summaries.

Impact

Found duplicates, ~10% of websites with errors, and outdated information through 10k+ providers. Nonprofits now use it to save hours determining which websites to prioritize.

Technology

10x faster deep research agent, semantic change detection, CSV processing, and headless browser automation with screenshot verification.

Sprint 1 Validation

Three Key Hypotheses Tested

Screenshots > Summaries: ✅ Validated - Screenshots save time verifying lack of hallucinations & seen as valuable compared to URL/text sourcing

Website Monitoring Killer Use Case: ✅ Validated - Nonprofits struggling to keep up to date with social service referrals

Nonprofit Payment Model: ❌ Invalidated - Not seen as high value enough to pay with limited resources, but extremely interested in using

Technical Innovation

Deep Research Agent

Built from-scratch agent 10x faster than OpenAI's by defaulting to looking for contact information on page, then scrolling conditionally alongside optimizations for finding contact info.

De-duplication Algorithm

Finds value counts of each variable, sends to LLM to identify stable vs typo variables, uses blocking to group similar entries, then batch processes differences.

Cost Efficiency

Used GPT-4.1-nano for cost and Claude Haiku for context. Total cost: $50 to identify 1k+ actionable insights on 5k websites vs $200 for competing solutions.

Key Learnings

Minimum Value Prop

Nonprofits cared more about analysis than direct control. Privacy implications of CSV uploads required approvals, so pivoted to recreating directories via public scraping.

Technical Challenges

Web is dynamic with shadow DOM, lazy loading, A/B tests. Built deterministic output system to determine real changes vs arbitrary differences.

User Focus

UI solving user problems > tech complexity. Checkbox verification system, crisp LLM explanations mattered more than fancy algorithms. Decision-makers over operators for adoption.

Origin Story

From Hackathon to Product

February 2024: Organized world's first AI Agents for Nonprofits hackathon

Recruited 8 collaborators from top tech companies, helped 2 nonprofits with data de-duplication

Became finalist at Anthropic/SPC event, built agent to grab missing email information

Connected with 3rd nonprofit in May, built full site with semantic flagging and screenshot proof