Mousey Housey

Landing Page
Listing Details Page
Graph showing listing that is fair valued.
Graph showing listing that is undervalued valued.
Graph showing listing that is fair valued.
Screenshot showing the Gemini summary of the listing prediction with easy to read listing factors.

Inspiration

Finding a “good deal” in housing is painfully opaque. Listing prices move with interest-rate shocks, local comps, and dozens of subtle signals buried in messy data. We wondered: what if one platform could read the market like Wall-Street quants, explain the numbers in plain English, and run anywhere from phone to quantum simulator? Mousey Housey was born at SpurHacks to do exactly that.

What it does

Pulls live market data (price, beds/baths, images, ZIP, macro rates).
Forecasts values 1-, 3- & 5-years out with a ZIP-aware LSTM ensemble.
Applies quantum Δ-correction to trim residual error on hard-to-predict ZIP codes.
Scores each listing (0-10) with Gemini and writes a one-paragraph investment memo.
Tags status — Undervalued, Fair, Overvalued — and shows the rationale.
Frontend candy: responsive cards, AI memo sheet, dark mode, rolling news ticker, animated price charts.

How we built it

Data layer (hackathon version) – CSV/JSON flat files committed to the repo for rapid iteration.
Scraper – Playwright scripts gather publicly available listing details with polite rate-limits and user-agent rotation.
ML pipeline
Pre-processing in Python (pandas, scikit-learn) with strictly causal lags & rolling stats.
TensorFlow/Keras LSTM models (128-64 units) trained per horizon.
PennyLane variational circuit (2 qubits, 9 learnable angles) wrapped in PyTorch for Δ-correction.
AI explanation layer – Gemini prompt → “score” + narrative.
Backend – Flask REST API in Docker, caches inference results for <1 s latency.
Frontend – Next.js 15, shadcn/ui, Tailwind, Recharts, Framer Motion.

(Road-map: swap flat files for a managed MongoDB Atlas cluster so we can ingest millions of listings and serve multi-tenant dashboards.)

Challenges we ran into

Non-leaky time-series feature engineering for hundreds of ZIP codes.
Stabilising quantum gradients (we scaled Δ by 10 k).
Playwright anti-bot evasion & 800 ms adaptive rate-limiting.
Streaming LLM prompts without blocking Next.js rendering.

Accomplishments that we're proud of

Trained multiple quantitative finance based housing price prediction models within a 0.82% MAPE average
Figured out an innovative solution combining proprietary housing prediction and quantum regression to forecast the price of an individual house (no other housing platform does this)
End-to-end demo in 36 hours (data → AI → UI).

What we learned

Training accurate price prediction models where there is close to 0 room for error.
Tiny quantum circuits can still deliver measurable gain when paired with classic DL.
LLMs excel at turning raw predictions into user-friendly advice — but prompt-stream handling matters.
Strict causal splits + ZIP-aware validation prevent data-leakage mirages.
Rate-limited, polite scraping keeps us within TOS and avoids IP bans.

What's next for Mousey Housey

Fine-tune Gemini with user feedback loops for even crisper advice.
Migrate data layer to MongoDB Atlas for real-time ingestion & scaling.
Train more models related to other factors that influence housing e.g mortgage rates, interest rates.
Deploy the quantum regressor on a real QPU (IonQ / IBM Q).
Launch Investor Pro: portfolio risk dashboard, PDF export, automated alerts.