Inspiration

Buying a home is often a family's biggest purchase, yet the information is scattered
Listings hide the real cost: stamp duty, registration, GST and deposits
Builders sell on super built-up area, not the carpet area you actually get
Builder agreements carry clauses most buyers cannot judge
The AI tools we found mostly stop at search or chat, so we wanted one that finishes the work

What it does

Mandate is an AI deal agent for Indian home buyers. The buyer gives one goal, like "2BHK in Whitefield under ₹60L, verify, book Saturday", and Mandate:

Plans the steps and searches listings
Re-plans on its own when the budget or a visit slot fails
Ranks homes on true cost, computed by code
Scans the builder agreement for RERA and clause risks, with questions to ask
Finds a free slot and waits for the buyer's approval before booking
Scores the lead silently and shows it live on the broker's dashboard
Shows every step in a live trace, tagged code or AI
Replies in English, Hindi, Kannada and more

Carpet-area loading, one of the checks, shows how much of the area you pay for is shared space, not usable home:

$$\text{Loading} = \frac{\text{Super built-up area} - \text{Carpet area}}{\text{Super built-up area}}$$

How we plan to build it

Mandate is at the design stage. This is our planned build.

Frontend: Next.js, Tailwind CSS, shadcn/ui
Agent: Vercel AI SDK with tool calling and Gemini Flash
Six steps: search, true-cost engine, contract scan, slot check, approval-gated booking, silent lead scoring
Safe math: costs, yields and scores in TypeScript, with zod-checked inputs
Data and live sync: Supabase with Realtime for the broker dashboard
Guardrails: tool order enforced in code, and no booking without the buyer's click
Demo data: seeded Bangalore listings and a synthetic builder agreement

Challenges we ran into

These are the challenges we expect, based on our design review.

Making recovery reliable when a search or slot fails
Stopping the agent from skipping steps, so we enforce tool order in code
Earning trust, so numbers come from code and every step is visible
Avoiding legal advice, so the scan flags clauses and questions only
Keeping Indian tax and RERA figures accurate, so rates live in one config file and are verified
Fitting everything into 24 hours, so we set checkpoints and a cut order

Accomplishments we are proud of

Studied existing real estate AI tools and projects to learn what works and where they fall short
Found the common gaps in what we reviewed:
- Most stop at search or chat and don't finish the deal
- Few explain the true cost after stamp duty, GST and deposits
- Few check the builder agreement for risky clauses
- Few show their steps or ask for approval before acting
- Few connect the buyer's chat to a live broker lead
Stress-tested the design and fixed the main flaws before writing code
Built human approval and a visible trace into the design from the start
Made it India-specific instead of generic