Predicate Ventures
All projects

Deployedengagement

Policy Edge AI

Compliance AI architecture for a govtech platform.

Worked with founder Christina Heintze Seavers to turn scattered regulatory documentation into an AI architecture and an execution plan. Built the prototypes that retired the critical unknowns early, then engineered the governed platform underneath. That work now reaches two regulated end clients through Policy Edge AI: a mortgage lender and a federal agency.

  • Governed per client
  • Human-approved changes
  • Keyless CI deploys
  • Prototype before build
  • Regulated & high-security environments
  • NLP & information extraction
  • Ontologies & knowledge graphs
  • Data science & evaluation
  • Cloud infrastructure

The problem

The product rested on a compliance-decisioning workflow: each inquiry needed a decision, with supporting evidence, against a structured body of regulation. A wrong call was costly, and each of hundreds of weekly inquiries took thirty to sixty minutes of analyst time. A small team faced a growing backlog. The alternative on the table was a six-figure monthly retainer, on a twelve-month minimum, from an enterprise consultancy.

The approach

Prototype the unknowns before committing to the architecture. Establish what the model could and could not decide reliably, then design the system around that boundary instead of assuming it away. The same rule carried into production. The machine proposes and a named human rules, and no accuracy figure goes out before a held-out test supports it.

How Policy Edge AI fits together

  1. Client systems

    • Mortgage lender: lakehouse and refinance model
    • Federal agency: audit-evidence system
  2. Per-client accounts

    Account Factory for Terraform, one account per client

  3. Security baseline

    Threat detection, CIS baseline, key rotation, human-approved changes

  4. Governed AWS organization

    Control Tower

Two client systems, each in its own account, on one governed organization.

What the Policy Edge AI work delivered

  • Prototype the unknowns first

    Establish what the model can and can't decide reliably before committing to the architecture around it.

  • The machine proposes, a human rules

    A named human decides every time, and no accuracy figure goes out before a held-out test supports it.

  • One governed account per client

    A flat AWS organization rebuilt as governed accounts under Control Tower, one account per client.

  • Structural controls, not policy

    Governance changes go through a human-approved plan, approve, apply pipeline; application deploys use keyless CI credentials.

What was built for Policy Edge AI

  • A flat AWS organization rebuilt as governed accounts under Control Tower and Account Factory for Terraform, one account per client
  • Threat detection, a CIS configuration baseline and key rotation, applied as code across every account
  • Governance changes go through a human-approved plan, approve, apply pipeline; application deploys use keyless CI credentials
  • Tamper-proof backups with verified restores, and a secure upload path with malware scanning and a quarantine step that fails closed
  • Policy Edge's AI-use, data-retention, customer-access and availability policies, authored and owned by Predicate
Blake came into a messy situation and quickly turned a big vision and scattered documentation into a platform we could actually deliver. By building the right prototypes early, he helped us answer the biggest open questions and move forward with confidence. He brought clarity and momentum exactly when we needed it.
Christina Heintze Seavers, Founder & CEO, Policy Edge AI

A mortgage lender: who is likely to refinance away

A mortgage lender is a Policy Edge AI client, reached through that platform rather than as a direct engagement. The question is retention: which borrowers in a mortgage servicing book are likely to refinance away, and roughly when.

Predicate moved the lender's loan system data off hand-built ETL and onto a governed lakehouse with scheduled ingestion. The lakehouse runs on Dagster, dbt, Athena and FastAPI on Fargate.

The lender's own system does not record whether a borrower later refinanced elsewhere. So Predicate built a pipeline over public property records that tells a refinance apart from a sale. It then trained a refinance model on those labels.

The model serves calibrated per-loan probabilities, uncertainty intervals and observed outcomes to the retention call list in the lender's app. It shipped behind a pre-registered deployment gate and was scored on rolling backtests with strict time cutoffs. Around ten feature experiments that failed to beat it are written up as negative results rather than quietly dropped.

  1. Loan system

    The lender's loan data

  2. Lakehouse

    Scheduled ingestion into governed layers

  3. Refinance labels

    Public property records separate a refinance from a sale

  4. Refinance model

    Calibrated probabilities with intervals

  5. Call list (a person decides)

    Staff queue loans for outreach

From the loan system to a call list, with refinance outcomes taken from public records.A person decides here

A federal agency: audit evidence, short of a verdict

Policy Edge AI delivers an audit-evidence system to a federal agency. Predicate engineered it; the contract, the audit methodology and every ruling sit with Policy Edge and its audit specialist.

During a financial-systems audit, the agency's finance staff answer auditors' document requests. The system picks up what they upload on a schedule and extracts text from PDFs, Office files, email and archives. Extraction runs in an isolated worker with no network access. The system then proposes the controls each request tests and compiles evidence packs for human reviewers.

Experts accept or correct each machine proposal in a review console that shows the machine's guess but never pre-fills it. The system reports how ready the evidence is and never issues an audit finding.

Predicate built the production platform as code: the API, the pipeline, the extraction worker and model inference, all inside the compliance boundary. Each model has a least-privilege role, records are append-only, and an evaluation harness applies pre-registered pass bars and held-out splits. It is live in a restricted deployment ahead of general availability.

The earlier prototype's accuracy figures were audited, found to be inflated by leakage, and withdrawn. No accuracy number goes out until the expert-labelled test sets are countersigned.

  1. Uploads

    Answers to auditors’ requests

  2. Extraction

    In an isolated worker with no network access

  3. Proposals

    Controls each request tests

  4. Expert review (a person decides)

    Accept or correct; never pre-filled

  5. Evidence pack

    Readiness, never a finding

  6. Reviewers (a person decides)

    The judgement stays theirs

The machine assembles and proposes. People decide, at two separate points.A person decides here

Where Policy Edge AI stands today

  • The mortgage lender: the refinance model is live behind the lender's retention call list
  • The federal agency: the audit-evidence system is live in a restricted deployment ahead of general availability
  • Both clients run in their own accounts on the one governed AWS organization

The Policy Edge AI tech stack

  • Python
  • FastAPI
  • Dagster
  • dbt
  • PostgreSQL
  • Docling
  • Amazon Bedrock
  • React
  • Next.js
  • Terraform
  • AWS Control Tower
  • Account Factory for Terraform
  • ECS Fargate
  • Aurora
  • Athena
  • Cognito