Applied enterprise AI · Washington DC
I build agentic systems that hold the reasoning of senior practitioners after the practitioners move on: geopolitical risk monitoring, policy tracking, crisis simulation, reputation intelligence. Not prototypes. Systems running for global organizations making high-stakes decisions.
Twenty-one years at APCO Worldwide, the last four spent architecting Margy, the firm's adaptive intelligence platform, with a team of engineers, data scientists and analysts across DC, New York and MENA.
Thesis
Standard operating procedures capture what an organization knows. They do not capture how its best people decide.
Succession planning treats expertise as a document problem. It is not. A thirty-year practitioner holds two different kinds of knowledge, and only one of them writes down.
Stays in the file cabinet
Leaves with the person
The tacit column is where the judgment lives, and it is the column that walks out the door. It cannot be transcribed, but a usable part of it can be elicited from the expert, encoded into agent architectures, and tested against decisions that practitioner actually made. That is the work. Not a replacement for the practitioner, and not a chatbot trained on their memos: a structure that reasons the way they reason, available after they are gone.
Organizations do not fail at this because of the technology. They fail because they never stopped treating it as a feature request.
Practice
Multi-agent orchestration, RAG pipelines, embedding pipelines and tool use. Agent identity design, guardrails and evaluation harnesses that constrain what each agent can claim, do and access, tested continuously rather than assumed.
Mental models extracted from human domain experts and encoded into sub-agent architectures, dashboards and decision-support frameworks. Variable-fidelity simulacra of senior expertise: systems that reason like the practitioners they extend, closer to simulated colleagues than chat tools.
Vendor diligence, disclosure standards and monitoring protocols designed into deployment rather than bolted on afterward, so regulated clients can adopt without regret.
AI strategy for leadership teams, model evaluation benchmarked against production requirements rather than vendor loyalty, and upskilling across technical and business functions.
Selected work
Core platform · APCO
APCO’s proprietary adaptive intelligence platform, and the concept the rest of the work specializes. Multi-agent systems for geopolitical signal, regulatory risk and stakeholder analysis, each build calibrated to one organization’s decision culture rather than shipped from a template. Delivered as $500K–$1.2M engagements across pharma, financial services, technology and food.
Specializations built on Margy
Continuous geopolitical and sector risk monitoring on licensed and proprietary data pipelines. Now in its third generation, with technology and defense capability tracks.
The same architecture tuned to health policy, regulatory and reimbursement signal, where the taxonomy and the escalation thresholds are different from the general build.
Built for the pharmaceutical operating environment: clinical and regulatory milestones, payer and access dynamics, and the stakeholder landscape around a product through its lifecycle.
Lab · APCO
Production work on applied AI for communications and advisory: distilled intelligence systems built from the knowledge of human domain experts.
Method
Mental model elicitation is the process of capturing, externalizing, and mapping how a human subject internally represents reality, handles deductive reasoning, or understands a complex system.
Full-lifecycle work in each case: technical discovery, agentic architecture, data pipelines, deployment and governance. Client names withheld.
Impact
The economics land in the first year. The capability compounds after that, and the second part is the one worth buying.
Direct cost
$300K+
Annualized
From retiring redundant tools, not from adding one more subscription.
Refocused talent
80%+
Of manual effort
Research, monitoring and content production automated. The analyst moves to interpretation, not out.
Decision speed
50%+
Faster
Detection, escalation and response, measured against the prior reporting cycle.
Speed to market
2×
On critical launches
Once the decision bottleneck is removed rather than worked around.
How it is built
Why it can exceed its sources
The three modes are not a metaphor. They are the conditions under which a model trained on many experts provably outperforms each of them, and they hold only when the expert set is heterogeneous enough. Abreu, Zhang, Malach and Saphra, A Taxonomy of Transcendence, COLM 2025. Which is the argument for interviewing twelve practitioners instead of cloning one.
Figures are indicative ranges observed across deployments, not results from any single engagement. Client names, engagement terms and deployment details are withheld.
Record
Twenty years watching technology move through a global firm. The web developer, the creative director, the strategist. Each rung is the reason the next one works, and the reason I can tell which part of an expert’s judgment is portable.
2022 →
Director, AI Solutions
APCO Worldwide
2018 → 2022
Director, Generative AI & Data Science
APCO Worldwide
2014 → 2021
Associate Creative Director, AI/ML Strategist
APCO Worldwide
2007 → 2014
Digital Art Director
APCO Worldwide
2005 → 2007
Web Developer
APCO Worldwide
2001 → 2004
Senior Technology Manager
Arteaga & Arteaga, San Juan
1998 → 1999
IT Group Intern
Microsoft, Puerto Rico
Also: Director of Data Science at DataScienceArts, brand consultant at HyperBrand Worldwide, and a board member at Maryland Gear.
Stack
Frontier models benchmarked against production requirements, not vendor loyalty. Whatever ships Tuesday is worth stress-testing by Friday.
Tooling
Disciplines
Credentials
Executive Certificates in three of MIT Sloan’s four tracks — Management and Leadership, Technology and Operations, Digital Business — earned across nine programs between 2018 and 2023.
2018
Artificial Intelligence: Implications for Business Strategy
Executive Certificate · MIT Sloan + MIT CSAIL
Sep 2018 → Dec 2022
Management and Leadership
Executive Certificate · track
Oct → Dec 2020
Machine Learning in Business
Management Certificate · MIT Sloan + MIT CSAIL
Aug → Oct 2021
Applied Business Analytics
Management Certificate
Aug 2021 → Oct 2023
Technology and Operations
Executive Certificate · track
Nov → Dec 2022
Business Implications of Extended Reality (XR)
Management Certificate
Dec 2022 → Sep 2023
Digital Business
Executive Certificate · track
Jul → Sep 2023
Data Monetization Strategy
Executive Certificate
Aug → Sep 2023
Artificial Intelligence in Pharma and Biotech
Executive Certificate
Degree and other programs
1995 → 1999
BA, Information Systems
Universidad Metropolitana, San Juan · IEEE
Oct 2021 → Feb 2022
Applied Data Science
MIT Professional Education · 93.64%
Aug → Oct 2020
Mathematics for Machine Learning
Imperial College London · three-course specialization
2018
Introduction to Artificial Intelligence
edX · 94%
Spanish and English, full professional.
Futures
The agents built for this work ingest global event and behavioral data, measure the actions of governments, markets, competitors and publics as they land, and forecast the paths those actions open: political and geopolitical risk on one side, the outcomes an organization wants on the other. The measurement exists so that leadership can see a path early enough to change it.
Simulation (no real events)Cities are index positions on the ring, not claims about themTen actions drawn from a pool of forty each cycle
What I build the agents to do well
Contact
If you are deciding what to build, what to buy, or whether a system will survive contact with real data and the people who have to use it, that is a good conversation to have.