Title AI

Title AI is the first platform that can autonomously search any US county recorder website and produce a complete title commitment report disrupting $3.6 trillion real estate industry.

Comment

Title AI is the first platform that can autonomously search any US county recorder website and produce a complete title commitment report disrupting $3.6 trillion real estate industry.

Inspiration

Real estate is a $3.6 trillion annual market in the US alone. Every single transaction - every home purchase, every refinance, every commercial deal - is blocked by the same bottleneck: the title search.

Before you can close on a house, a human examiner must manually log into county government websites, trace ownership history back decades, and hunt for liens, judgments, and encumbrances that could void the sale. This process takes 5 to 14 business days. It costs $300–$600 per search. And it hasn't fundamentally changed in over 50 years.

Why hasn't anyone automated it?

Because there are 3,600 counties in the United States, and not a single one offers a standardized API. Every county has its own website. Its own forms. Its own search logic. JavaScript-heavy interfaces, session state, CAPTCHA-like flows, pagination quirks. The only way to access this data is the same way a human does it: through a web browser.

No scraper survives this. No RPA tool handles the variation. No traditional automation can adapt to 3,600 different government UI designs that change without warning.

Amazon Nova Act can. And that's why we built Title AI.

This isn't a demo. This isn't a proof of concept navigating one website. This is a production-grade platform that deploys AI browser agents against real US county recorder websites, extracts real property records, and produces industry-standard title reports - doing in minutes what takes human examiners weeks.

We're not optimizing one step of the real estate transaction. We're eliminating the biggest bottleneck in the entire industry.

What it does

Title AI is the first platform that can autonomously search any US county recorder website and produce a complete title commitment report - the document that every real estate transaction in America requires before closing.

Enter a property address. Six AI agents execute immediately:

  1. County Resolution Agent - resolves any US address to the correct county recorder website from a verified database of 110+ counties across 22 states, with US Census geocoder fallback and city-to-county mapping for 100+ cities

  2. Nova Act Browser Agent - provisions a cloud Chromium session via AgentCore Browser Tool, navigates the live county recorder website, fills out search forms, paginates through results, and extracts deed records and parcel history. On the actual government website. In real time. With every screenshot streaming live to the user's screen so they can watch the AI work.

  3. Chain of Title Agent (Nova Pro) - traces every ownership transfer: grantor, grantee, document type, recording date, instrument number - returned as typed structured data, not free text

  4. Lien Detection Agent (Nova Pro) - identifies mortgages, tax liens, HOA liens, judgments, and encumbrances - structured, auditable, with source citations

  5. Risk Assessment Agent (Nova Pro) - classifies every title exception by severity (Fatal / Curable / Informational), flags chain-of-title gaps, and generates actionable recommendations

  6. Report Agent (Nova Pro) - generates an executive summary and a branded ALTA-compliant Title Commitment Report PDF following American Land Title Association 2021 standards, with Schedule A (property details, vesting, legal description) and Schedule B (requirements and exceptions)

Every data point traces back to its source. Every ownership record, every lien, every exception carries a provenance citation - source URL, retrieval timestamp, document excerpt, confidence score. This isn't a black box. It's an auditable intelligence system.

The real-time experience is unlike anything in real estate technology: you watch an AI browser agent navigate a government website frame-by-frame, filling forms and extracting records, while analysis agents process the results in parallel. A process that takes a human examiner days plays out on your screen in minutes.

Why this matters beyond title search

Title search is the critical path of real estate. It's the reason closings take 30–45 days instead of 30 minutes. When you can automate title search:

  • Instant closings become possible - the 30-day closing timeline exists primarily because of title and escrow. Remove that bottleneck and same-day closings become real.
  • $27B title insurance industry transforms - underwriting costs plummet when structured, auditable AI reports replace manual examination
  • Real estate becomes liquid - properties can be verified, cleared, and transferred at the speed of digital transactions, not the speed of county paperwork
  • Every adjacent industry accelerates - mortgage lending, refinancing, property investment, REITs, iBuying - all are gated by title verification

We're not building a title search tool. We're building the infrastructure that makes real estate move at the speed of the internet.

How we built it

Nova Act - the breakthrough that makes this possible

The core insight: county recorder websites are the last mile of real estate data, and the only way through that last mile is a browser controlled by an AI that can reason about unfamiliar interfaces.

Nova Act runs on an EC2 sidecar connected to AgentCore Browser Tool, which provisions cloud-hosted Chromium sessions with live view URLs and CDP WebSocket access. The agent navigates each county's unique website - filling search forms, handling pagination, clicking through deed records - while streaming screenshots at every step.

We built county-specific prompt strategies for each recorder's unique interface: Travis County's search requires different form fields than Cook County's, Maricopa County paginates differently than Harris County. Nova Act adapts to each. For unseen counties, a generic fallback prompt handles the variation - Nova Act's reasoning capabilities let it navigate forms it's never seen before.

Nova Pro - structured intelligence pipeline

Four Nova Pro agents process the raw documents that Nova Act extracts. Every agent uses generateObject() with typed Zod schemas - structured JSON output, not free text. Chain of title returns OwnershipNode[]. Lien detection returns Lien[]. Risk assessment returns TitleException[]. The report agent synthesizes everything into an ALTA-compliant format.

This is not prompt engineering and prayer. Structured output means the system cannot produce malformed data. The UI always renders. The PDF always generates. The audit trail is always complete.

Real-time screenshot streaming

Nova Act's browser work runs on a background thread. Screenshots flow into a queue. An SSE generator polls the queue and yields each frame to the client in real time. Users and judges see the AI working through government websites as it happens - navigating, searching, extracting - not a loading spinner and a result, but the actual work being done.

Conversational interface with tool calling

A Nova Pro chat agent exposes run_title_search and get_search_report as callable tools. Users can initiate searches, ask follow-up questions about results, and review reports entirely through natural language. The agent has context of recent search history and can reason about results across multiple properties.

Voice interface (Nova Sonic)

Nova Sonic powers a speech-to-speech voice interface - title professionals can initiate searches and review results hands-free while doing fieldwork.

Durable execution

Searches run via Inngest with per-step retries, real-time progress updates, and 10-minute timeouts. Each step - county lookup, browser extraction, chain analysis, lien detection, risk assessment, report generation - updates a Turso (libSQL) database with percentage complete, logs, and screenshots. If a step fails, it retries. If a county site is unreachable, fallback agents take over.

Full stack

Layer Technology
Browser Automation Amazon Nova Act + AgentCore Browser Tool (cloud Chromium)
Intelligence Amazon Nova Pro (us.amazon.nova-pro-v1:0) - 4 analysis agents + chat
Voice Amazon Nova Sonic - speech-to-speech interface
Embeddings Amazon Titan Embed v2
Document Processing AWS Textract - PDF text extraction with FORMS + TABLES
Model Hosting Amazon Bedrock
Compute EC2 (Nova Act sidecar) + Vercel (Next.js frontend)
Framework Next.js 16 + Vercel AI SDK v6 + TypeScript
Database Turso (libSQL) - search history, job state, reviews
Background Jobs Inngest - durable execution with per-step retries
Auth Clerk
PDF Generation jsPDF - ALTA-compliant branded reports

Challenges we ran into

3,600 different government websites, zero standardization. Every county recorder site has different form fields, different search flows, different result layouts. Some require parcel numbers before showing deed history. Some use JavaScript-heavy interfaces that destroy traditional scrapers. We had to build a verified county database with tested navigation flows - and engineer Nova Act prompts that are specific enough to succeed on known counties while general enough to adapt to unknown ones.

Streaming screenshots from a blocking browser agent. Nova Act's browser execution is synchronous. We architected a background-thread system with a queue.Queue as the communication bridge, letting the SSE generator stream screenshots to the client while the agent continues working - giving users real-time visibility into a process that would otherwise be a black box.

Sequential AI analysis on top of an already-slow browser step. Four Nova Pro agents must process the extracted documents. We parallelized chain-of-title and lien detection since they operate on the same raw data independently - cutting the analysis phase roughly in half while maintaining structured output guarantees.

ALTA compliance is not optional. Title commitment reports follow a specific industry standard (American Land Title Association 2021 Commitment Form). Schedule A, Schedule B Part I, Schedule B Part II - each has required fields and a required format. We built a 565-line PDF generation module that produces reports indistinguishable from those created by human title companies.

Accomplishments that we're proud of

  • First-ever autonomous title search on live US government websites - Nova Act navigates real county recorder portals, extracts real property records, and produces real title reports. No mocks. No cached data. No simulations.
  • 110+ counties across 22 states with verified recorder URLs and tested navigation flows - and growing
  • ALTA-compliant Title Commitment Reports - industry-standard PDFs with Schedule A/B, confidence badges, risk banners, and full source citations
  • Complete provenance chain - every data point traces to its source with URL, timestamp, excerpt, and confidence score. Every report is fully auditable.
  • Real-time screenshot streaming - users watch Nova Act navigate government websites frame-by-frame as it works, turning an opaque process into a transparent one
  • Structured output guarantee - every Nova Pro agent uses typed schemas. The system cannot produce malformed data. Zero parsing failures.
  • Human-in-the-loop review workflow - section-level commenting, approval/rejection, revision requests - because AI-generated title reports need human oversight before underwriting decisions
  • 5–14 days of manual work replaced by minutes - the fundamental value proposition of the platform

What we learned

Nova Act is the first technology we've seen that can survive government websites. County recorder portals are the final boss of web automation - inconsistent, JavaScript-heavy, session-dependent, updated without warning. Nova Act handles them because it reasons about the UI rather than following brittle selectors. The critical skill: writing prompts at the right abstraction level - specific enough to find the target data, general enough to adapt to layout variation.

Structured output changes everything for production AI pipelines. generateObject() with typed schemas eliminates an entire class of runtime failures. When Nova Pro returns OwnershipNode[] instead of a paragraph of text, the UI always renders, the PDF always generates, and the audit trail is always intact. For any team building production systems on foundation models: structured output is not a nice-to-have. It's the difference between a demo and a product.

The real estate industry is desperate for this. Every title professional we spoke to had the same reaction: "If this actually works on live county sites, this changes everything." The technology is ready. The market need is undeniable. The only question is coverage - and that's an engineering problem, not a research problem.

What's next

This is not a hackathon project that ends at the hackathon. Title AI addresses a $27 billion market with a working product.

  • Scale to all 3,600 US counties - the architecture handles any county. Expansion is a URL curation and validation problem, not an engineering redesign. We're adding counties weekly.
  • Continuous title monitoring- run automated searches on a schedule, alert property owners and lenders when new liens, ownership changes, or encumbrances appear. Turn point-in-time searches into real-time property intelligence.
  • Title insurance underwriting integration - feed structured Nova Pro output directly into underwriting platforms. When the input is structured, typed, and auditable, underwriting decisions can be partially automated.
  • Instant closing pipeline - partner with escrow companies, lenders, and real estate platforms to enable same-day closings. Title search is the bottleneck. We remove it.
  • International expansion - land registries in the UK, Australia, Canada, and parts of the EU are online but require manual browser navigation. The same Nova Act architecture applies globally.
  • Commercial real estate - more complex ownership structures, easements, CC&Rs, zoning records, environmental liens. Higher value per transaction. Same core platform.

Title AI. The bottleneck is gone.

Built With

Share this project:

Updates