SMART DECISION ENGINE
Product Development
My Debt Plan
6 months
Product Lead, Algorithm Design
Intelligent recommendation system that analyzes complex financial data points to match users with optimal debt relief products. Built using multi-criteria logic evaluation and weighted scoring.
Project Overview
In the financial services industry, matching a customer to the right product requires analyzing dozens of variables—income, debt levels, creditor profiles, and historical data. Manual matching is slow and prone to error. The Decision Engine automates this process, providing instant, accurate recommendations.
The Challenge
- 01 Data Complexity - processing 50+ variables per user profile
- 02 Real-time Performance - matching must happen in under 1 second
- 03 Regulatory Compliance - ensuring logic follows industry guidelines
- 04 Accuracy - matching must be more accurate than manual advisor review
- 05 Transparency - providing clear reasons for each recommendation
Logic & Algorithm
The engine uses a tiered filtering and scoring approach:
- 1 Hard Filters - Immediate exclusion based on non-negotiable criteria (e.g., minimum debt level)
- 2 Weighted Scoring - Assigning points to positive and negative indicators for each available product
- 3 Creditor Preference Analysis - Factoring in the likelihood of specific creditors accepting different product types
- 4 Risk Assessment - Evaluating the long-term sustainability of the recommended solution
- 5 Final Ranking - Providing the top 3 options with percentage match scores
Technical Implementation
Custom Python-based evaluation engine built for high-concurrency processing.
FastAPI framework ensuring sub-second response times for frontend applications.
Key Outcomes
<1s
85%
Lessons Learned
Building a decision engine isn't just a technical task—it's a knowledge engineering task. The most critical part of the project was the weeks spent with subject matter experts to translate their intuition into mathematical weights.
Edge cases will always surprise you. A robust engine needs to handle missing or conflicting data points without failing, often by defaulting to "human required" flags when confidence scores are low.