Compliance monitoring · Production

Zillow builds a Fair Housing Guardrails system for LLM-powered real estate conversations

The problem

Out-of-the-box LLMs applied to real estate conversations could violate Fair Housing Act requirements by steering users based on legally protected characteristics, creating legal risk and inequitable service at scale.

First attempt

Prompt-based compliance instruction produced high recall but non-deterministic, low-precision behavior that over-flagged permissible queries. Stop-list matching lacked contextual awareness, misclassifying location names and accessibility-related queries as non-compliant.

Workflow diagram · grounded in source
1
User submits real estate query
trigger
“conversational interface that allows Zillow users to discuss real-estate topics with the objectives of understanding the process, identifying their desired location and type of home”
2
Stop list syntactic check
validation
“parse the input query, looking for terms that match the stop list, using syntactic matching”
3
BERT classifier detects FHA violations
ai_action
“implemented a classification model based on Bidirectional Encoder Representations from Transformers (BERT), a now well-known technique introduced by Devlin et al. (2019). We fine-tuned the base sequence classification model with a binary…”
4
Post-processing guardrail review
validation
“in its role as a post-processing component, the fair housing guardrail reviews the output from the LLM, flagging any content that might potentially violate fair housing regulations before it is displayed to the user”
5
Flagged content routed to predefined message
routing
“For flagged content, a predefined message is displayed instead of the LLM output”
6
Human review and iterative improvement
feedback_loop
“periodically sampling data for human review allows for the identification of false positives and false negatives. This not only helps in rectifying any immediate inaccuracies by updating the stoplist component, but also contributes addit…”
Reported outcome

Zillow designed and deployed internally a combined Fair Housing Guardrails system uniting prompt engineering, a stop list, and a fine-tuned BERT sequence classifier to detect FHA violations both pre- and post-LLM processing, with iterative improvement via human review and feedback.

Reported metrics
Recall level at maximum precision lift0.6
Prompt-based approach precision qualitysub-optimal precision, inadvertently withholding permissible information
Reported stack
BERTLLM
Source
https://www.zillow.com/tech/navigating-fair-housing-guardrails-in-llms/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Zillow designed and deployed internally a combined Fair Housing Guardrails system uniting prompt engineering, a stop list, and a fine-tuned BERT sequence classifier to detect FHA violations both pre- and post-LLM proc…

What tools did this team use?

BERT, LLM.

What results were reported?

Recall level at maximum precision lift: 0.6; Prompt-based approach precision quality: sub-optimal precision, inadvertently withholding permissible information (source-reported, not independently verified).

What failed first in this deployment?

Prompt-based compliance instruction produced high recall but non-deterministic, low-precision behavior that over-flagged permissible queries.

How is this compliance monitoring AI workflow structured?

User submits real estate query → Stop list syntactic check → BERT classifier detects FHA violations → Post-processing guardrail review → Flagged content routed to predefined message → Human review and iterative improvement.

Related compliance monitoring cases