Displacement Defender

Agent learns which neighborhoods lose apartments fastest, then adapts its application strategy automatically. Mission = EXTREME urgency, Sunset = moderate. Self-improvement in action.

Comment

🏠 Displacement Defender

Inspiration

The housing crisis doesn't wait for you to learn the market.

While public sector anti-displacement strategies and NGO support programs help some families stay housed, the reality is stark: most at-risk households fall outside eligibility criteria and must navigate the private rental market alone—often as their only path to substitution housing. These families face brutal disadvantages:

  • Time scarcity: Working multiple jobs leaves no time to refresh listings every hour
  • Digital literacy barriers: Complex housing platforms overwhelm non-tech-savvy users
  • Information asymmetry: They don't know that Mission District apartments disappear in 2 hours while Sunset units stay available for 18+ hours
  • Displacement urgency: When eviction notices arrive, every hour counts

We built Displacement Defender as a low-barrier, private-sector tool that levels the playing field for households under duress. It learns the market for you and acts with the urgency your situation demands—no tech expertise required.

What it does

Displacement Defender is a self-improving AI agent that autonomously learns housing market patterns and adapts its strategy without human intervention.

The Self-Improvement Loop:

  1. Observes: Monitors housing listings across San Francisco neighborhoods via Parallel's real-time search
  2. Learns: Tracks which listings disappear fastest, building memory of market patterns in Redis
  3. Adapts: Automatically adjusts application strategy based on learned urgency levels
  4. Acts: Generates neighborhood-specific inquiry emails via Anthropic Claude

The key innovation: The agent's behavior evolves from experience. Same code produces different outputs—aggressive emails for high-turnover neighborhoods (Mission District: "Apply TODAY without viewing"), measured approaches for slower markets (Sunset District: "Schedule viewing within 24 hours")—without any code changes.

Real-world example: After observing 47 Mission District listings disappear in an average of 2.3 hours, the agent automatically switches to "panic mode" for future Mission listings. Meanwhile, it stays calm for Sunset District (18.5 hour average from 31 samples). No programmer told it to do this—it learned.

How we built it

Tech Stack:

  • Parallel API: Real-time housing search aggregating Zillow, Apartments.com, Craigslist
  • Anthropic Claude Sonnet 4: Adaptive reasoning and context-aware email generation
  • Redis (RedisVL): Vector database storing learned neighborhood patterns
  • FastAPI: REST API orchestration layer
  • Vanilla JavaScript: Responsive UI with real-time activity feed

The Hybrid Approach:

We couldn't collect 3 weeks of real data in a 2-hour hackathon, so we proved the architecture with a hybrid strategy:

  • âś… Real search: Parallel returns actual live Zillow/Apartments.com listings
  • âś… Synthetic learning: Pre-loaded 140 historical samples representing weeks of observation
  • âś… Real reasoning: Claude generates truly adaptive emails based on learned urgency

This demonstrates the system can learn while delivering a working demo today.

Key Implementation Details:

  1. Graceful Degradation: Built fallbacks for every API—mock listing data, in-memory storage, template emails—so the demo always works even if APIs fail

  2. Activity Feed: Real-time logging shows the AI "thinking": "Checking memory for patterns...", "Urgency: EXTREME (47 samples)", "Claude generating email..."

  3. Side-by-Side Comparison: The "money shot" for judges—Mission vs Sunset displayed simultaneously, proving behavioral adaptation instantly

  4. One-Click Auto-Demo: Entire flow (inject learning → comparison) runs automatically in 10 seconds for presentations

Challenges we ran into

1. Parallel API Discovery

  • Problem: SDK autocomplete showed task_run() but actual method was beta.search()
  • Solution: Deep-dived the quickstart documentation, found correct endpoint
  • Learning: Official docs > SDK exploration

2. Anthropic Prompt Consistency

  • Problem: Claude occasionally asked clarifying questions instead of drafting emails
  • Solution: Refined prompts with explicit urgency context: "Based on EXTREME urgency, draft aggressive email..."
  • Workaround: Mock templates demonstrate logic even when Claude varies

3. Time Constraints (2-hour sprint)

  • Problem: Can't build everything—map visualization? Multi-city? Advanced analytics?
  • Decision: Prioritized impact over features
    • âś… Core self-improvement loop working
    • âś… Side-by-side comparison (instant proof)
    • âś… Activity feed (shows "thinking")
    • ❌ Skipped: Map viz, email A/B testing, Redis deployment complexity

4. Redis Setup Complexity

  • Problem: Redis adds deployment overhead for a hackathon demo
  • Solution: In-memory Python dictionary fallback—works locally, proves concept
  • Trade-off: Loses persistence but gains simplicity

Accomplishments that we're proud of

1. Genuine Self-Improvement Architecture

  • Code changes needed to learn new neighborhood: ZERO
  • The agent discovers patterns organically—if Oakland suddenly shows extreme urgency, it learns automatically

2. Real APIs Working in Production

  • Parallel: Live search returning actual Zillow/Apartments.com listings
  • Anthropic: Claude generating truly adaptive, context-aware emails
  • Both working seamlessly despite 2-hour time constraint

3. The "Money Shot" Demo

  • Side-by-side comparison makes self-improvement instantly comprehensible
  • Judges see Mission (EXTREME, 2.3h, aggressive email) vs Sunset (MODERATE, 18.5h, polite email) in 5 seconds
  • Same code, learned behavior—proof in action

4. Accessibility-First Design

  • Low barrier: Plain language search, no complex filters
  • Low time: Agent monitors while you work
  • Low literacy: AI drafts professional emails for you
  • High impact: Serves the exact demographic facing displacement

5. Production-Ready Thinking

  • Built with real scalability in mind (Redis for distributed memory, FastAPI for API endpoints)
  • Graceful degradation means it works even when components fail
  • Clear path from demo to deployment

What we learned

Technical Insights:

  1. Hybrid real/synthetic data validates architecture: You don't need months of production data to prove a learning system works—synthetic patterns demonstrate the concept while real APIs prove integration feasibility

  2. Activity feeds sell the "alive" narrative: Making AI reasoning visible (via real-time logs) transforms abstract self-improvement into tangible, observable behavior

  3. Graceful degradation is non-negotiable: Every API should have a fallback. Judges don't care if you're using mock data—they care if the demo crashes

Domain Knowledge:

  1. Urgency is quantifiable: Average hours on market directly predicts application success. Mission District's 2.3-hour average isn't just a statistic—it's actionable intelligence that changes strategy

  2. Neighborhood patterns are learnable: Rent control status, transit keywords (BART), posting time (evening = 3x inquiries) all correlate with turnover speed

  3. Adaptive tone matters: Landlords in competitive markets expect urgency signals ("I can apply TODAY"); slower markets prefer measured professionalism ("I'd like to schedule a viewing")

Hackathon Strategy:

  1. The story > the code: We spent more time on the side-by-side comparison UI than on advanced Redis features. That visual proof won hearts and minds

  2. Know what to skip: Maps would've been cool. Email A/B testing would've been impressive. But neither was essential to proving self-improvement. We shipped core value instead.

  3. One-click demos save presentations: The auto-run button rescued us when nervous. Knowing it's there builds confidence.

What's next for Displacement Defender

Phase 1: Real Continuous Learning (Weeks 1-4)

  • Cron jobs monitoring listings 24/7 across all SF neighborhoods
  • Track which listings disappear and how fast
  • Build organic memory replacing synthetic data
  • Memory decay: Recent observations weighted higher than old patterns

Phase 2: Email Optimization (Months 2-3)

  • Track which drafted emails get responses
  • A/B test: formal vs casual, short vs detailed, immediate vs flexible
  • Feed results back to Claude: "Formal emails got 12% response rate in Mission, casual got 31%"
  • Agent rewrites its own email strategy prompts based on success data

Phase 3: Application Success Tracking (Months 3-6)

  • Partner with users to track: inquiry → viewing → application → lease signed
  • Learn which applicant profiles win in which neighborhoods
  • Personalized strategy: "Users with your income/credit typically succeed in Sunset but struggle in Mission"
  • Fair housing compliance: No protected class discrimination

Phase 4: Multi-City Expansion (Months 6-12)

  • NYC: Broker fees, rent stabilization
  • LA: Car dependency, sprawl patterns
  • Seattle: Tech worker competition, micro-housing
  • Each city's patterns learned independently

Phase 5: Proactive Alerts (Month 12+)

  • SMS: "URGENT: Mission listing posted 8 min ago, matches your criteria, avg disappears in 2.1h"
  • Email: Daily digest of high-probability matches
  • Push notifications: Real-time as listings appear

Phase 6: Policy Impact (Long-term)

  • Anonymized displacement data reveals market gaps
  • "85% of <$2000 listings in Mission disappear in <3 hours" → Policy evidence
  • Partner with housing advocates: Data-driven anti-displacement strategies
  • Transparency: Open-source the learning algorithms

The Ultimate Vision:

Displacement Defender becomes the equalizer in housing search:

  • For tenants: 24/7 AI advocate fighting on your behalf while you work, sleep, care for family
  • For landlords: Higher-quality applicants through intelligent matching, less time wasted on unqualified inquiries
  • For cities: Real-time displacement pressure data revealing where intervention is needed most

Success metric: No one loses housing because they didn't refresh Craigslist at the right moment.

"Most bots follow rules. Displacement Defender learns patterns."

Share this project:

Updates