cfpb

HMDA mortgage lending disparity analyzer — denial rates, racial disparities, lending deserts, and lender benchmarking

Automatic Tagging of CFPB Complaints using LLM's and Ensemble Techniques

Audit of the CFPB Consumer Complaint Database as a source for measuring collections compliance. Two categories of FDCPA conduct its taxonomy cannot express, a silent date-parsing defect, and template-driven volume inflation.

How accurately an LLM sorts real CFPB complaints into 9 categories: 76.5% against an 11% baseline.

End-to-end CFPB complaint analytics using Python, DuckDB, SQL, machine learning, Power BI, automated tests, and GitHub Actions.

An interactive app that tells the story of the 2008 mortgage crisis and its decade-long aftermath through five borrowers, built on 64 million HMDA mortgage records (2007–2017) from the CFPB.

Classify thousands of rows of free text with an LLM: batching, cache, resume, USD budget guard, low-confidence review flags, and accuracy measured against real labels. Reports its own cost and accuracy instead of asserting them.

Analysis of 10.7M CFPB consumer complaints (2021-2025) with Python, SQL, and Tableau

Production-grade RAG system over CFPB consumer financial complaint narratives featuring FastAPI REST API, Streamlit UI, Chroma vector store, anti-hallucination guardrails, and Docker deployment.

Fine-tuned DistilBERT for CFPB consumer-complaint routing (8 classes, macro-F1 0.85), plus a quantitative faithfulness audit of its explainability layer: Integrated Gradients, SHAP and attention rollout agree on only 13-29% of their top-attributed tokens.

Early warning for conduct-risk incidents at banks from 4.7M real CFPB complaints. Kafka/Event Hubs → Delta Lake → negative-binomial detection with market adjustment → Snowflake/dbt → Power BI. Point-in-time backtest: 8/11 incidents, median 29 days ahead of monthly MIS.

Consumer complaint classifier fine-tuned on DistilBERT, with a TF-IDF baseline comparison and confidence-gated human-review routing. 0.87 weighted F1 on held-out test data.

Open-source PII detection and re-identification risk benchmarks for LLM pipelines under EU AI Act, HIPAA Safe Harbor and GLBA NPI.

Local SLM for financial complaint triage: Qwen3-4B QLoRA fine-tuned on 70K CFPB complaints, served via llama.cpp. 90.9% product accuracy, ~204 ms latency on RTX 4070, Streamlit dashboard, TF-IDF baseline.

RAG-based legal Q&A agent built with LangChain, ChromaDB, and the Anthropic Claude API — query the CFPB Supervision and Examination Manual using natural language with cited, grounded answers.

An escalation-calibrated support agent over 396,952 CFPB consumer complaints. The deliverable is a frontier: how much of the queue closes unattended, at each error rate you would accept.

NLP on CFPB consumer-complaint narratives for 10 banks/fintechs: TF-IDF + NMF topics, template detection, quarterly alt-data signals in DuckDB vs. stock returns

Streamlit ML app that classifies public CFPB complaint narratives into customer feedback themes.

LangChain, LlamaIndex and CrewAI tools for 24 public open-data Apify actors: sanctions, SEC Form D, FMCSA, permits, jobs, domains. Maintained by the actors' author; pay-per-result.

Early-warning system for emerging consumer scams — clusters public CFPB/FTC complaint data to surface novel fraud patterns before they hit banks' books.