Mortgage Policy & Guideline Agent — Sanity Challenge Path One

Amol — Sat, 19 Sep 2026 15:16:15 +0000

This is a submission for the Sanity Challenge, Path One: Ship an Agent That Queries Real Content

What I Built

I built a Mortgage Policy & Guideline Agent that answers mortgage eligibility and policy questions using structured mortgage content stored in Sanity.

Mortgage policy questions can look simple but often depend on several connected pieces of information: loan-product limits, underwriting guidelines, exceptions, and the source supporting each rule.

Instead of relying only on an LLM's memory, I created a structured Sanity knowledge base containing:

Mortgage Policies
Guidelines
Loan Products
Exceptions
Sources and citations

The agent can retrieve the relevant structured content before answering a question.

For example, a conventional mortgage decision can depend on:

Minimum credit score
Maximum DTI
Maximum LTV
Loan purpose
Occupancy
Applicable guideline
Documented exceptions
Exception conditions
Supporting source

The goal is to make the answer traceable back to structured content rather than simply producing a plausible AI response.

Demo

Live Application

Mortgage Policy & Guideline Agent

The companion web application provides a simple interface for asking mortgage questions and viewing:

Eligibility results
Important rules
Borrower and product facts
Source citations
Sanity-backed responses

Try These Prompts

1. Eligibility

Conventional purchase, 640 credit score, 47% DTI and 95% LTV. Is the borrower eligible? Explain the rules that determine the answer.

2. Product Comparison

Compare the Harborstone 30-year and 15-year conventional products. What are the differences in minimum credit score, maximum DTI and maximum LTV?

3. Exception Reasoning

A conventional borrower has a DTI above the standard limit. What documented exceptions could apply, and what conditions must be satisfied?

4. Source Verification

What source and effective date support the maximum DTI rule for the Harborstone conventional product?

5. Scenario Change

A borrower has 700 credit, 47% DTI and 95% LTV. Does the answer change compared with a borrower with 640 credit? Explain why.

6. Explain the Decision

Don't just tell me whether this borrower qualifies. Walk through credit score, DTI, LTV, applicable guideline, exception eligibility and supporting source.

The Harborstone mortgage catalog is fictional demo data created for this project and is not official lending guidance.

Code

GitHub — Mortgage Policy & Guideline Agent

The repository contains the Sanity schemas, structured demo content, GROQ queries, agent instructions, MCP configuration, and companion web application.

How I Used Sanity

The core of the project is the structured mortgage knowledge base in Sanity.

I modeled mortgage information as connected content rather than a single collection of unstructured documents.

The main content types are:

Loan Products

Products contain structured underwriting limits such as:

Minimum credit score
Maximum DTI
Maximum LTV
Minimum down payment
Loan type
Occupancy
Loan purpose
Product code

Guidelines

Guidelines contain structured thresholds such as:

Metric
Operator
Value
Unit

For example, a guideline can represent a DTI threshold such as DTI <= 45%.

Exceptions

Exceptions contain:

Conditions
Compensating factors
Approval requirements
Exception identifiers
Relationships to applicable guidelines

Sources

Sources provide the supporting information used for citations, including:

Source title
Published/effective date
Locator
Excerpt
URL where applicable

This structure allows the agent to retrieve the information needed to answer a question instead of treating mortgage policy as plain text.

Sanity Context MCP

I connected my Cursor AI agent to Sanity Context MCP for the Path One implementation.

The agent uses Sanity Context MCP to query the live Sanity dataset and retrieve the structured mortgage content.

The flow is:

User question

↓

Cursor AI agent

↓

Sanity Context MCP

↓

Sanity Project krg1w0kt / production

↓

Structured mortgage content

↓

Product + Guideline + Exception + Source

↓

Grounded answer with supporting information

The MCP tools used by the agent include document querying and retrieval through the Sanity Context MCP connection.

I also created project-specific Cursor rules and a mortgage-policy-agent skill that instruct the agent to use Sanity content when answering mortgage policy questions and to provide citations.

Why structured content matters

Consider this question:

Conventional purchase, 640 credit score, 47% DTI and 95% LTV. Is the borrower eligible?

The agent needs more than a keyword match.

It needs to identify the relevant conventional product, retrieve the applicable DTI and credit thresholds, check the borrower's values against those thresholds, and determine whether a documented exception applies.

That is where the structured Sanity content becomes important.

The same Sanity dataset is also used by the companion web application, which provides a user-friendly interface for the mortgage agent.

Sanity Project Details

Project ID: krg1w0kt

Dataset: production

The Sanity project contains the structured mortgage knowledge base used by the agent.

The primary document types are:

Mortgage Policy
Guideline
Loan Product
Exception
Source

The project ID is provided so the Sanity team can inspect how the content was modeled and used.

Agent Session

The Path One agent was developed and tested in Cursor using Sanity Context MCP.

An agent-session transcript is optional for this submission. If included, it will be curated to show the relevant parts of the development process and Sanity MCP interaction.

Any published transcript will be checked for API keys, tokens, credentials, or other sensitive information before publication.

sanitychallenge

Fetch - scan your dog, then chat, plan, and hear their story

Amol — Sun, 16 Aug 2026 10:30:00 +0000

This is a submission for Weekend Challenge: Dog Days Edition

What I Built

Fetch is an AI dog companion that starts with a single photo and turns it into a small toolkit for looking after your dog.

You upload a photo, and Fetch identifies the most likely breed along with temperament, size, energy level, and a few care tips. From there, everything stays connected to that same dog:

A chat assistant that already knows the breed, so the advice you get is specific to your dog instead of generic. After a scan it also suggests breed-specific questions about grooming, feeding, exercise, environment, training, and health.
A one-tap care plan covering feeding, grooming, exercise, environment, training, and vet care, which you can read in the app, download as a PDF, or print.
A storybook that writes a short story starring your dog and reads it aloud in a natural voice.

The idea was to make something that feels less like four separate tools and more like one companion that actually remembers your dog across the whole app. I wanted a new dog owner to be able to point their camera at their pup and immediately have somewhere to ask questions and get a plan.

Demo

Live app: https://fetch-dog-ai.vercel.app/

Try it by uploading a clear photo of a dog on the Scan tab, then move through Ask Fetch, Care Plan, and Storybook. The storybook narration plays right in the browser.

Code

https://github.com/amolbhandari52/fetch-dog-ai

How I Built It

The frontend is a single HTML file with plain JavaScript and CSS. No framework, no build step, so it loads fast and works the same on a phone as on a laptop. The scanned dog's profile is held in a small shared state object, which is what lets the chat, care plan, and storybook all stay aware of the same dog.

All the AI calls go through serverless functions on Vercel so that the API keys stay on the server and never reach the browser. There are two functions: one for Google Gemini and one for ElevenLabs.

A couple of decisions I am happy with:

Rather than hard-code a Gemini model name, the function asks the API which models the key can use and picks a current one. I ran into a model being retired mid-build, so I made it self-correct instead of breaking. It also retries with backoff and falls back to another model when Gemini returns a temporary high-demand error, which kept the care plan and storybook reliable during testing.

For structured output like the breed profile and the care plan, I use Gemini's JSON response mode so the app can render the results directly instead of parsing loose text.

The storybook narration is generated per page and streamed back as audio. The ElevenLabs function auto-selects an available voice on the account, so it works without me hard-coding a voice ID.

Best Use of Google AI - Gemini handles four separate jobs: vision breed identification, the breed-aware chat, the structured care plan, and the story generation, all through one endpoint.

Best Use of ElevenLabs - the storybook narration is powered by ElevenLabs text to speech, which is what turns the generated story into something you actually listen to.

DEV Community: Amol

Mortgage Policy & Guideline Agent — Sanity Challenge Path One

What I Built

Demo

Live Application

Try These Prompts

Code

How I Used Sanity

Loan Products

Guidelines

Exceptions

Sources

Sanity Context MCP

Why structured content matters

Sanity Project Details

Agent Session

sanitychallenge

Fetch - scan your dog, then chat, plan, and hear their story

What I Built

Demo

Code

How I Built It