CivicLens

Interactive Map
landing
Usecase

Inspiration

A small business owner expanding from a food truck into a brick and mortar restaurant faces the exact same regulatory complexity as a national chain (permits, zoning, health code, alcohol licensing), but without the in house legal team to navigate it.

We observed that:

Small business owners are expected to interpret the same regulatory landscape as large enterprises, with none of the legal counsel
Regulatory information is scattered across municipal, county, state, and agency sources with no unified source of truth
Generic LLM chatbots happily generate plausible sounding compliance answers with zero source grounding, which is a serious liability in a legal or regulatory context
Expansion decisions (new city, new license type, new service line) compound complexity exponentially, turning what should be a quick lookup into hours of manual cross referencing

We decided to build CivicLens around one non negotiable rule: if a compliance claim can't be traced to a verifiable government source, it doesn't get shown to the user. Everything else, including the RAG pipeline, the dashboard, and the Diff Engine, was built in service of that one rule.

What It Does

A user describes their business in plain language:

"I own a food truck in Dallas and want to open a restaurant in Austin with a beer garden."

CivicLens classifies the business, retrieves the relevant regulations from a curated knowledge base, and generates a personalized compliance dashboard, showing required permits, risks, and compliance gaps, each with a direct link to the underlying government source.

Capability	Description	Impact
Natural Language Business Classification	Parses a free text business description into structured business type, location, and service attributes	Removes the need for users to know regulatory taxonomy up front
Regulatory Retrieval (RAG)	Vector search over a curated, embedded corpus of TX regulatory documents	Surfaces requirements relevant to this business, not generic guidance
Citation Validation Layer	Rejects any compliance finding that lacks a verifiable source URL	Eliminates hallucinated requirements before they reach the user
Regulatory Diff Engine	Side by side comparison of requirement changes between two business scenarios	Answers "what changes if I move or expand" instantly
Compliance Pulse	Ongoing monitoring of new ordinances, regulations, and legislation	Prevents owners from being blindsided by regulatory changes
Personalized Compliance Dashboard	Aggregates findings, risks, and source links into a single view	One place to act instead of a pile of government PDFs

Snapshot of the build:

[x] 1 fully validated end to end expansion scenario (food truck to restaurant, Dallas to Austin)
[x] Citation enforced RAG pipeline over a curated Texas regulatory corpus (TABC, city ordinances, state licensing guides)
[x] pgvector backed semantic search, no separate vector database required
[x] Structured, citation required prompting pattern enforced at the application layer
[x] Built in 11 days, validated through 15+ full demo rehearsals

==Every compliance finding shown to a user carries a verified source link. No exceptions.==

System Architecture

Layer	Technology	Role
Frontend	Next.js 14, React, Tailwind CSS	Business intake form, compliance dashboard, Diff Engine UI
API Layer	NestJS (Express)	Orchestrates classification, retrieval, validation, and response
Database & Search	PostgreSQL + pgvector	Stores regulatory document embeddings, semantic search with no separate vector store
Data Pipeline	Python, LangChain	Ingests, chunks, and embeds public regulatory sources
Embeddings	OpenAI `text-embedding-3-small`	Converts regulatory text chunks into searchable vectors
Reasoning & Generation	Claude (structured, citation required prompt)	Synthesizes retrieved context into findings, never freeform

Keeping pgvector inside the primary Postgres instance, rather than standing up a separate vector database, was a deliberate trade off for the 11 day build window. One less moving part is one less thing that could break during demo week.

How We Built It

Phase 1: Problem Scoping & Knowledge Base Definition

Rather than attempting to cover all of Texas, we picked a narrow, deeply validated slice: a small set of highly regulated business types (food service, alcohol service) in a single expansion scenario, food truck to restaurant, Dallas to Austin. Primary sources identified for ingestion: the Texas Licenses & Permits Guide, TABC requirements, relevant city ordinances, and recent legislation.

Phase 2: RAG Pipeline Construction

A Python and LangChain ingestion pipeline chunks regulatory documents, embeds them with OpenAI's text-embedding-3-small, and stores them in pgvector for semantic retrieval.

// Example: compliance finding object, as enforced by the citation validation layer
{
  id: "F-RESTAURANT-001",
  category: "Health & Safety Permit",
  description: "New permit required when transitioning from mobile food
                service to a fixed restaurant location",
  citation: {
    source: "Texas Licenses & Permits Guide",
    url: "<verified-government-source-url>"
  },
  status: "validated"
}

Every finding generated downstream must conform to this shape. No citation.url, no display to the user.

Phase 3: Compliance Analysis & Citation Enforcement

Retrieved regulatory context is passed to Claude using a structured prompt that requires a citation for every finding. A validation layer sits between model output and the dashboard:

        Claude Output
              |
              v
   Citation Validation Layer
              |
       +------+------+
       v             v
  Valid source    No source
  URL present     URL found
       |             |
       v             v
  Shown to user   Rejected,
                  never shown

Phase 4: Regulatory Diff Engine

The signature feature. Rather than generating diffs on the fly and risking hallucinated deltas, the Diff Engine runs on pre validated expansion scenarios, food truck to restaurant being the primary one, with manually verified requirements on both sides of the comparison.

 Scenario A (Current)        Scenario B (Target)
    Food Truck                   Restaurant
    Dallas                       Austin
        |                            |
        +-------------+--------------+
                       v
            Regulatory Diff Engine
                       |
                       v
          New Requirements + Risk Deltas

Phase 5: Demo Hardening & Documentation

Every citation included in the demo was manually checked against its original government source. We ran 15+ full practice runs to surface edge cases ahead of time.

Full System Flow

Business Description
 ("I own a food truck and want to open a restaurant")
            |
            v
 Business Classification
            |
            v
 Regulatory Retrieval
 (pgvector Search + RAG)
            |
            v
 Citation Validation
            |
            v
 Compliance Analysis
 & Risk Detection
            |
            v
 Regulatory Diff Engine
            |
            v
 Personalized Dashboard
 & Actionable Guidance

Data Flow & Validation

Flow	Path	Validation / Protection
F1	User input to Business Classifier	Input parsing, business type normalization
F2	Classifier to pgvector Search	Embedding based relevance threshold on retrieval
F3	Retrieved context to Claude (structured prompt)	Generation constrained to retrieved context only
F4	Claude output to Citation Validation Layer	Source URL required, unsourced claims rejected
F5	Validated findings to Diff Engine	Only pre validated, manually verified scenarios are diffed
F6	Findings to Dashboard	Manual spot check during MVP, direct links to gov sources

Challenges We Ran Into

Challenge 1: RAG Hallucinations

The Problem: Large language models are confident even when they're wrong, and in a compliance product, a confidently wrong answer is worse than no answer at all.

Our Approach: A citation validation layer sits between model output and the user. Any finding without a valid, verifiable source is automatically rejected.

Result: Zero unsourced compliance claims reach the dashboard. The trade off is coverage, not correctness.

Challenge 2: Regulatory Complexity

The Problem: Requirements vary significantly across industries, cities, and agencies, with no clean, unified schema to query against.

Our Approach: We narrowed to a small set of highly regulated business types and municipalities, building a scalable architecture while keeping accuracy high within scope.

Result: A narrow but trustworthy product, with an architecture designed to widen later without a rebuild.

Challenge 3: Time Constraints (11 Days)

The Problem: Eleven days is not enough time to build broad regulatory coverage and a polished, reliable demo.

Our Approach: We prioritized aggressively. The RAG pipeline and one fully polished scenario took precedence over geographic breadth or additional business categories.

Result: One scenario that works end to end and holds up under scrutiny, instead of many that don't.

Challenge 4: Accuracy Verification

The Problem: A single incorrect permit requirement is enough to undermine trust in the entire platform.

Our Approach: Every citation included in the demo was manually checked against its original government source before being shown.

Result: Time consuming, but non negotiable, and the reason the demo could survive hard questioning.

Accomplishments That We're Proud Of

1. Citation Enforced Regulatory Intelligence, Built in 11 Days

Most teams exploring regulatory AI end up with generic chatbots that produce plausible but unreliable answers. CivicLens was designed from day one to prioritize verifiable information over coverage.

2. The Regulatory Diff Engine

Technically straightforward, but it answers a question business owners constantly face: "what changes if I move or expand?" It delivers, in seconds, guidance that normally takes hours of legal research.

3. Accuracy & Trust Risk Matrix

Reframing our core challenges as a risk register helped us prioritize where engineering effort actually mattered:

Risk	Likelihood	Impact	Mitigation
Unsourced or hallucinated compliance claim	High (default LLM behavior)	Critical (legal/financial harm to user)	Citation validation layer rejects unsourced findings
Geographic or industry scope gaps	High (broad regulatory landscape)	Medium	Deliberately narrowed MVP scope to TX, limited business types
Outdated regulation shown to user	Medium	High	Compliance Pulse monitoring (in progress)
Conflicting information across government sources	Medium	Medium	Manual spot check against original source during MVP

4. Demo Reliability Through Repetition

15+ full practice runs let us identify edge cases ahead of time and walk into the final demo with confidence instead of hope.

5. Grounding in a Real Business Owner's Story

Anchoring the entire presentation in a real small business journey, food truck to restaurant, turned an abstract regulatory problem into something a judge, or any small business owner, could immediately feel.

What We Learned

1. LLMs Are Powerful, But High Stakes Domains Require Guardrails

The instinct to let a model answer everything quickly breaks down in legal and regulatory contexts. Source grounding, citation validation, and human review aren't optional polish. They are the product.

2. Regulatory Data Is Messier Than Expected

Government websites use inconsistent terminology, update on different schedules, and occasionally contradict each other. Working against real government data taught us things a clean dataset never could.

3. Focus Wins Hackathons

We deliberately said no to multi state coverage, predictive legislation features, and user accounts. Those "no"s are what let us invest in accuracy, polish, and a memorable demo instead of spreading thin.

4. Technology Alone Doesn't Solve Civic Problems

The real value comes from understanding the underlying problem, choosing the right technical approach, and designing for trust, not from picking the flashiest model.

What's Next for CivicLens

Phase 2: Near Term Implementation

Priority	Item	Timeline	Impact
P0	Compliance Pulse email delivery	2-3 weeks	Closes the loop on ongoing regulatory monitoring
P0	User authentication + saved profiles	1-2 weeks	Enables persistent, returning user dashboards
P1	Expand Diff Engine beyond food truck to restaurant	4-6 weeks	Generalizes the signature feature across scenarios
P1	Broaden Texas business category coverage	1-2 months	Moves past the narrow MVP scope

Phase 3: Monetization & Coverage

Priority	Item	Timeline	Impact
P1	Launch subscription tier ($9.99 to $19.99/mo)	1-2 months	Establishes a sustainable monetization path
P2	Multi location support	2-3 months	Serves owners operating across multiple jurisdictions
P2	Advanced compliance insights	2-3 months	Deepens value beyond basic permit lookup

Phase 4: Distribution & Partnerships

Partner with Small Business Development Centers, chambers of commerce, and local economic development organizations to offer CivicLens as a member benefit
Position partnerships as the primary distribution channel for reaching small business owners who wouldn't otherwise find a tool like this

Phase 5: National Scale & Platform Integration

Expand regulatory coverage nationally beyond Texas
Integrate with accounting and point of sale platforms via API, embedding CivicLens into existing business workflows
Launch a mobile experience

Key Targets, End of Q4 2026

[ ] 100% of compliance findings remain citation validated: zero unsourced claims shown to users, no exceptions
[ ] Regulatory Diff Engine covers 5+ fully verified expansion or relocation scenarios, up from 1 at MVP
[ ] Compliance Pulse live for an initial pilot cohort of small businesses
[ ] At least one Small Business Development Center or chamber of commerce partnership in active discussion