Inspiration:

The Law Office Study Foundation’s challenge: help residents and businesses while minimizing liability and burden on city government. McKinsey’s November 2026 State of the States survey found that learning what’s required is one of the most consequential steps in dealing with government, and Texas launched an AI regulatory assistant (TREO) this year. Glendale’s small-business owners, many of them not native English speakers, have no city-level equivalent. Inspiration

What it does

BizPath gives one verified set of Glendale requirements two ways. Lifecycle chips (start, change location, hire, received a notice, renew, find help) show only what applies now. Business type pages show everything in order, readable without AI. Owners can ask a question in their own language and get a cited answer, paste a City notice to have it explained, and print a visit sheet listing what to bring. Every requirement links to its source and code section and shows why it applies. City staff get a live dashboard: answers with a verified source, legal-advice requests declined, hand-offs, content gaps, rules people find hard to understand, and entries due for review.

How we built it

Next.js on Vercel, with the interface generated in v0 and the logic built with Claude Code. A plain rules table decides which requirements apply. Claude Haiku 5.5, through the AI SDK, only explains and translates those entries. Upstash Redis stores questions and feedback with no personal data.

Challenges we ran into

Keeping the model from adding plausible details that weren’t in the source. We fixed it by moving every decision into the rules and letting the model write only summaries.

Accomplishments that we're proud of

  • The AI never decides the law. A plain, readable rules table picks which requirements apply. The model only explains and translates them, and every fact carries its source, its code section and the date it was last verified.
  • It passed its own test. We ran a ten-question evaluation on the live site covering accuracy, citations and refusals. It answered in Armenian and Spanish with fees unchanged, declined a request for legal advice about workers’ tips, and said “I don’t have a verified source” instead of guessing about sidewalk dining.
  • It caught and fixed its own overreach. In testing, the model once added a step that wasn’t in the source and once chose between two legal outcomes for an expired certificate. We tightened the design so it now lists what the City’s page says and sends the owner to the right desk.
  • Two ways in, from one dataset. Lifecycle chips give owners only what applies to their situation; business type pages give the full picture and work with no AI at all.
  • A notice decoder. Paste a letter from the City and it identifies the rule, quotes the notice’s own words, links the law and names who to call.
  • A live staff dashboard built on real use. It shows answers backed by a verified source, legal-advice requests declined, hand-offs, content gaps, rules people find hard to understand and entries due for review, with no personal data stored.
  • Honest about its limits. Unverified items appear as links only, baseline pages are labelled as such, and every page says it is a prototype and not an official City website.
  • Built solo in one afternoon, deployed and working end to end.

What we learned

  • Where the AI sits matters more than which AI you use. Letting a model answer legal questions directly is what made New York’s business chatbot give illegal advice. Putting the decision in rules and using the model only to explain made the answers consistent and checkable.
  • A model will add plausible details unless the design stops it. Telling it “only use the sources” wasn’t enough. What worked was giving it less to do: the tables, contacts and disclaimers come from data, and the model writes only the summary and next steps.
  • “I don’t know, here’s who to call” is a feature. For a legal tool, a clear refusal and a hand-off to a named desk protect the owner and the City better than a confident guess. Many people would rather read than chat. McKinsey’s 2026 survey found most residents still prefer standard digital channels over AI ones, so the browse pages that need no AI are as important as the ask box.
  • Organize by the owner’s moment, not the agency. Starting, hiring and receiving a notice call for different subsets of the same rules, and owners think in those moments.
  • Feedback is content maintenance. Recording which rules people find hard to understand, and which questions have no verified source, tells a city exactly what to fix next.
  • Verification is the slow part. Writing the app took an afternoon; confirming each fee and rule against the official page is the real work, which is why every entry carries a date and unverified ones are shown as links only.

What's next for Glendale BizPath

  • Native-speaker review. Have Armenian and Spanish speakers check the interface text and the most common answers, then add reviewed Korean, Tagalog and Farsi.
  • Deeper coverage. Verify the type-specific requirements for retail, personal services, offices and hotels, and confirm the county and state steps that currently appear as links only.
  • A 90-day pilot with the City. Work with one content owner in Economic Development or the Permit Services Center to review entries monthly, and measure what the pilot is meant to test: whether owners arrive better prepared and whether routine questions at the counter go down.
  • More notice types. Extend the notice decoder beyond the three kinds it recognizes today, including photo upload.
  • Voice intake. Let owners describe their situation aloud in their own language, with the same rules engine deciding what applies.
  • Compliance reminders. Turn the deadlines view into optional renewal and filing reminders. A reviewed legal referral. Add a “Need a lawyer?” pointer to a legal help resource chosen with the Law Office Study Foundation.
  • A model other cities can reuse. The rules table and tagged corpus are plain files, so another city could adopt the same structure with its own requirements, and the published llms.txt helps general AI assistants find Glendale’s official pages.

Built With

  • anthropic
  • claude
  • claude-code
  • next.js
  • react
  • redis
  • shadcn-ui
  • tailwindcss
  • typescript
  • upstash
  • v0
  • vercel
  • vercel-ai-sdk
Share this project:

Updates

Submission history