FlatFair by Team Techbros

Inspiration

Buying a first home means deciding where your life will happen: how long you will commute, whether everyday essentials are nearby, and what you can afford without overstretching.

PropertyGuru’s H2 2023 Consumer Sentiment Study found that 51% of surveyed non-property owners in Singapore found securing their first home difficult. The survey covered property types broadly, but it captures the uncertainty facing people trying to take that first step.

For prospective HDB resale buyers, relevant evidence is spread across transaction records, maps, transport services and housing information. A household must piece it together while balancing different needs.

FlatFair asks: What if buyers could describe their situation and have an AI agent investigate realistic housing options—with evidence they can inspect?

What FlatFair will do

FlatFair is our proposed solution to C1: HDB resale market intelligence and affordability forecasting in the Databricks AI Social Impact Challenge.

It combines a conversational housing assistant with market analytics, fair-value estimates and six-month price forecasts.

Consider a couple looking for a four-room resale flat with a $650,000 budget. One works in Jurong East, the other in the CBD. Both want manageable commutes and nearby groceries.

FlatFair will help them:

Define their requirements. Confirm their budget ceiling, commute limits, preferred flat attributes and remaining-lease requirements.
Investigate suitable blocks. Retrieve comparable transactions, property records, routes and nearby amenities.
Compare an explainable shortlist. Show how each candidate fits their requirements, including each person’s commute and the compromises involved.
Assess an asking price. Compare a buyer-entered price with a modelled transaction-value range and supporting comparable sales.
Explore the outlook. View historical trends, six-month town-level forecasts and an affordability index based on clearly identified income benchmarks.
Inspect the evidence. See sources, observation dates, sample sizes and uncertainty behind the recommendation.

The shortlist will identify blocks worth investigating. Historical transaction data does not establish that a particular flat is currently available.

What makes the agent useful

The agent will do more than answer questions about a fixed dataset. It will plan an investigation, select relevant tools, inspect the returned evidence and decide whether another retrieval is needed.

For example, if a candidate block has too few recent comparable transactions, the agent can request a longer historical window within predefined limits. It must disclose the older evidence and flag any remaining uncertainty.

The buyer stays in control. Code will enforce hard constraints and evidence requirements, so the agent cannot silently raise a budget ceiling, relax a commute limit or invent missing information. Buyers can change their priorities and rerank the shortlist using the evidence already collected.

Our data

We plan to combine the challenge’s sources with additional public data that supports household-level decisions:

HDB resale transactions: comparable prices, historical trends and predictive-model training.
HDB Property Information: building attributes and block-level context.
HDB Annual Report 2025: housing supply and policy context.
SingStat population by planning area: aggregate demographic context.
SingStat household income statistics: consistent affordability benchmarks.
SLA OneMap: geocoding, routes and available amenity layers.
LTA DataMall bus-stop locations: additional transport context, subject to API access.

We will preserve source dates and definitions, distinguish straight-line distances from routed journeys, and keep national income benchmarks separate from a buyer’s personal budget.

How we plan to use Databricks

Databricks will support the pipeline from raw public data to buyer-facing recommendations.

Lakeflow Jobs and Delta tables will ingest and retain dated source snapshots. SQL and Python transformations will normalise records, handle missing values, flag outliers and quarantine incompatible schema changes.

MLflow will track experiments for two separate predictive models:

A six-month forecasting model, initially using gradient-boosted trees with historical monthly features, compared against last-value and seasonal-naive baselines.
A fair-value model using transaction-level attributes such as location, floor area, storey and remaining lease.

We will evaluate models using time-based holdouts, report prediction errors and interval coverage, and flag areas with insufficient evidence.

Unity Catalog will govern data access and lineage, with explicit source and freshness metadata.

Databricks Apps and Foundation Model APIs will power the conversational experience. The agent will use constrained SQL tools and external API adapters to retrieve evidence. Application code will enforce permissions, retrieval budgets and ranking rules.

AI/BI dashboards will provide interactive price trends and affordability analysis, with filters for town, flat type, storey range and year.

Suggested extensions

We are incorporating fair-value estimation and Unity Catalog governance into the proposed build.

Our stretch goal is BTO supply analysis: exploring how upcoming supply relates to nearby resale transactions. This requires verified, dated geospatial supply data beyond the annual report. We will present any findings as associations unless the evaluation supports a stronger conclusion.

Challenges we anticipate

The main challenges are joining records reliably across sources, comparing genuinely similar flats, and providing useful estimates where transactions are sparse.

We also need to prevent misleading precision. A price model cannot fully observe renovation condition, and a forecast cannot guarantee future prices. FlatFair will make these limitations visible alongside its outputs.

For the two-week sprint, we will prioritise a complete buyer journey, validate API access and Databricks Free Edition quotas early, and limit route enrichment to selected towns before expanding coverage.

How we will measure success

We will assess both technical reliability and usefulness to prospective buyers.

Technical evaluation will cover forecast performance against baselines, valuation error, uncertainty coverage, retrieval failures and unsupported agent claims.

For an initial pilot, we plan to recruit 10 prospective buyers and compare equivalent shortlisting tasks using FlatFair and a basic dashboard. Our exploratory targets are:

30% less median time to complete a shortlist.
At least 8 of 10 participants able to explain the main trade-offs behind their choices.

These are proposed targets, not measured results.

Current stage and what comes next

FlatFair is currently a Round 1 concept and architecture proposal. The buyer journey, data sources, model approach and evaluation plan are defined; implementation and validation are the next steps.

Our aim is to give prospective HDB buyers a clearer basis for deciding where to look, how to assess an asking price, and which compromises they are willing to make.