CollateralQC
AI-powered QC platform for real-estate appraisals
ML Engineer @ Syncd Analytics
The Problem
Real-estate appraisal management companies (AMCs) had to manually review hundreds of UAD XML appraisal reports against complex rule sets — a slow, error-prone process consuming 4–6 hours per report. Quality control was inconsistent across reviewers.
What I Built
CollateralQC is a full-stack AI platform that automates this entire workflow. I engineered a 194-rule deterministic QC engine (version 5.0.0) that parses MISMO 2.6/3.6 XML, evaluates compliance rules, and produces a pass/fail/N/A report in seconds. On top of that, I built a RAG-powered Doc Intelligence layer using ChromaDB so reviewers can query appraisal documents naturally.
Technical Deep Dive
The QC engine uses a YAML-defined rule set (TF, IFTHEN, IMAGE, DTD categories) with real MISMO XPath lookups for field extraction. Evidence search scans narrative fields like reconciliation summaries, income analysis, and valuation methods. DSPy powers a rule enhancement system that improves rule accuracy over time. The backend is Django REST + Flask microservices; the frontend is Next.js 14 with Tailwind and React Query. Deployed on DigitalOcean via Docker Compose, with Cloudflare R2 for document storage.
Impact
Reduced manual appraisal review time by 60–80%. Benchmark results on the 7416HEAT test file: PASS=99, FAIL=0, NA=65, VISION=30 — fully deterministic, zero false failures. AMC staff went from spending hours per report to reviewing a pre-scored result in minutes.