Advanced AI & Analytic Systems

Where intelligence becomes operational judgment.

Phronesis designs advanced AI, data science, and analytic systems for mission environments where the cost of error is high and clarity is non-negotiable.

Our Work

Active Signal

Cross-Border

Company Thesis

Phronesis builds safety-constrained AI and intelligence systems for high-stakes government missions where uncertainty, adversarial behavior, and governance matter as much as model performance.

Safety-Constrained AI

Systems designed with behavioral guardrails, policy layers, and rigorous evaluation from the ground up.

Adversarial Robustness

Models that hold under stress, paraphrase drift, and deliberate manipulation — not just clean inputs.

Governance by Design

Reproducible evaluation, audit trails, and policy-compliant architectures built for oversight.

Judgment Under Uncertainty

Calibrated outputs that signal confidence correctly — essential when ambiguity has operational consequences.

What We Do

Data science in service of consequential decisions.

Phronesis applies large language models, advanced data science, and safety-critical system design to problems where the stakes are high and conventional approaches fall short.

Our work sits at the intersection of model performance and operational judgment — building systems that are not only accurate, but interpretable, auditable, and safe to deploy in environments where failure has real consequences.

Large Language Model Research

Custom LLM development with behavioral safety constraints, policy-layer architecture, and reproducible red-team evaluation for high-stakes, adversarial environments.

Anti-Money Laundering

Advanced transaction network analysis, financial signal extraction, and machine learning pipelines for detecting and tracing illicit financial flows at scale.

Anti-Human Trafficking

Data-driven identification of trafficking networks using pattern recognition, behavioral signal modeling, and cross-jurisdictional data integration.

Evaluation & Governance

Structured rubrics, audit harnesses, and versioned test suites that make model behavior measurable, defensible, and suitable for independent review.

Our Principles

Intelligence that holds
under scrutiny.

Safety-Constrained

Every system we build has explicit behavioral constraints, refusal mechanisms, and policy layers — not as an afterthought, but as a design primitive. Safety is not a feature. It is the architecture.

Adversarially Robust

Our models are evaluated against deliberate manipulation, paraphrase drift, and adversarial stress testing — not just clean validation sets. Robustness is measured, not assumed.

Governance by Design

Reproducible evaluation, versioned audit logs, and policy-compliant architectures built for independent review. Every output is traceable, every decision is defensible.

Calibrated Uncertainty

In high-stakes environments, overconfidence is a failure mode. Our systems are designed to signal uncertainty correctly — because knowing what you don't know is operationally essential.

Transparent Tradecraft

We document our methods, publish our evaluation rubrics, and make our analytic logic legible to oversight. Transparency is the foundation of trust in mission-critical systems.

Human-Centered Oversight

AI augments analyst judgment — it does not replace it. Our systems are designed with human-in-the-loop review, escalation protocols, and clear boundaries on autonomous action.

How We Work

From raw data to
operational judgment.

Collect & Ingest

We build pipelines to ingest heterogeneous, high-volume data — regulatory filings, unstructured text, financial transactions, behavioral signals — and normalize it for downstream analysis.

Model & Constrain

Custom LLMs and probabilistic classifiers are trained, calibrated, and wrapped in policy layers that define allowed outputs, refusal behaviors, and escalation logic before any result is surfaced.

Red-Team & Evaluate

Every system undergoes adversarial stress testing against curated prompt suites, paraphrase drift, and ambiguity gradients. Failures are logged, classified, and mitigated before deployment.

Deliver & Document

Outputs are packaged with reproducibility artifacts — versioned datasets, model checkpoints, audit logs — suitable for independent technical review and government program reporting.

Featured Platform

Geofin

Deployed to the U.S. Department of Treasury · FinCEN

The standard of care for strategic Bank Secrecy Act data analysis and regulatory compliance. A map-based analytical platform for multi-resolution geographic exploration of financial regulatory filings — developed in partnership with DARPA and deployed to the Financial Crimes Enforcement Network.

Multi-Resolution Visual Analytics

Explore patterns in regulatory data at global, national, local, and entity scales. Customizable summary statistics and visualizations allow analysts to rapidly triage locations and surface trends across space and time.

Real-Time Aggregation

Analyze arbitrary subsets of data in real time. When datasets are too large for exact computation, the platform supports approximate aggregation with a configurable margin of error — delivering fast, actionable results at scale.

Data Augmentation

Import external datasets to contextualize primary data and reduce noise from extraneous factors. Enables analysts to layer their own intelligence on top of regulatory filings for richer, more defensible analysis.

Secure Authentication & Audit Logging

FinCEN-specific modifications include secure authentication, complete query logging for audit purposes, and role-based access controls — meeting federal security and compliance requirements out of the box.

Custom Entity Resolution

BSA filings present inconsistent and untrustworthy location data. Geofin's in-application entity resolution interface lets administrators resolve and propagate geotag corrections across the full dataset — eliminating duplicated analyst effort.

→

Built for FinCEN Analysts

Through sustained engagement with Treasury analysts, Geofin's features are continuously refined to simplify tasks that previously required high technical competence — lowering the barrier to sophisticated financial intelligence analysis.

Leadership

Guided by research,
oriented toward mission.

Casey King

Founder & Chief Executive Officer

Casey King is a data scientist and technologist with deep expertise in large language models, machine learning systems, and the application of big data analytics to complex policy and governance problems. He leads technical strategy at the intersection of data infrastructure and applied AI.

Casey has consulted for federal, research, and national-security customers and is recognized for applying large-scale analytical systems to high-stakes problems — including counter-terrorist financing, anti-money laundering, and anti-human trafficking initiatives. He has taught data science and policy courses at Yale and delivered keynote addresses at major analytical forums including Battle of the Quants.

At Phronesis, Casey brings enterprise-scale technical leadership and rigorous research methodology to the design of safety-constrained AI systems for government and national-security missions.

Role

Founder & CEO

Education

MS, UC Berkeley · PhD, Yale

Affiliations

Yale University

Focus

AI Safety · Data Science · Policy