Propseller delivers its seller and buyer clients the best outcomes by combining top in-house salaried agents, a team of 100 specialists, software and AI. Our agents close 26x more transactions than the average Singapore agent. This success has made Propseller the most awarded agency in Singapore, with 92% of our agents recognized at the SEAA Awards 2024. With over 1,000 five-star reviews and an average rating of 4.8/5, property sellers and buyers trust Propseller to handle their property transaction.

---

My Contributions

A. Data Infrastructure

• Built a property knowledge graph linking:
  • Transactional data (URA, SRX, 10M+ past listings)
  • Location features (walkability scores, MRT/bus stops, school districts)
  • Building attributes (floor level, facing, renovation quality via CV)
• Implemented automated data pipelines (Airflow + AWS Glue)

B. Model Development

• Hybrid Approach:
  • Gradient Boosted Trees (XGBoost) for baseline valuation
  • Neural Additive Models (NAMs) for interpretable feature weighting
  • Ensemble Layer to reconcile predictions
• Key Innovations:
  • Incorporated time decay weighting (recent sales > older comps)
  • Added neighborhood momentum indicators (price trends by postal code)

C. Deployment & Validation

• C. Deployment & Validation
  • AI suggestions vs. human appraisers (n=2,300 properties)
  • Result: 2.1% mean absolute error vs. final sale prices
• Explainability:
  • SHAP values showing top price drivers (e.g., “MRT <500m adds 4.2%”)
  • Agent override analysis (tracked when pros disagreed with AI)

Impact:

80% faster valuations (now <2 hours)
15% reduction in price overestimation
Integrated into Propseller’s agent workflow dashboard

---

2. Conversational AI System

Problem Statement

• High lead drop-off due to slow response times
• Agents wasted 30% time on repetitive queries

My Contributions

A. Core Components

1. Intent Recognition
   • Fine-tuned DistilBERT on:
     • 50K+ labeled chat logs
     • 12 core intents (viewings, financing, paperwork)
   • Achieved 93% accuracy
2. Knowledge Base
   • Vectorized 500+ documents (HDB rules, CPF policies)
   • Used FAISS indexing for instant retrieval
3. Response Generation
   • Hybrid Rule-Based + LLM:
     • Pre-approved templates for common queries
     • GPT-3.5 for nuanced responses (fine-tuned on top agent replies)
   • Guardrails:
     • Confidence thresholding (<85% ? human escalation)
     • Toxicity filtering

B. Continuous Learning

• Implemented agent feedback loop:
  • “Thumbs up/down” on AI responses
  • Retrained biweekly on corrected answers

Impact:

24/7 query handling
40% reduction in agent workload
28% improvement in lead conversion