Wuhao Xia

AI strategy · Implementation · Applied analytics

I design and implement AI systems for complex business and policy workflows.

I work from opportunity identification and process redesign through architecture, hands-on build, validation, and human review—connecting AI strategy to working systems.

Review selected work Download résumé

Experience across World Bank Group, Bain & Company, and Capgemini

Columbia SIPA · MPA candidate, Data Science for Policy · May 2027

Selected work · 02–04

Three complementary ways I turn analysis into a decision tool.

Product design, failure-aware AI architecture, and time-aware model validation. Each case is organized around a user decision and the evidence needed to support it.

02Decision system

Grid Intelligence

An investigation layer for fragmented U.S. grid data.

I built the data pipeline, analytical logic, cloud layer, and interface for users who need to connect operational risk, market signals, and transition conditions without proprietary market-intelligence tools.

Built
Public-data ETL, BigQuery analytical layer, six connected interface modules
Key judgment
Surface contradictions for investigation instead of hiding them inside one recommendation score
Case study Public code
EXECUTIVE BRIEFING / PJM
Investigation Briefing Anomalies Arbitrage Transition
Cross-source signal

Transition potential is not matched by queue activity.

Demand conditions
Interchange pattern
Transition conditions
Investigate

Compare peer conditions and source coverage before forming a recommendation.

Reconstructed from the implemented interface structure; no operational recommendation is implied.

Design judgment: explain contradictions instead of smoothing them away.
03Applied AI design

SEC 10-K RAG

Risk analysis that separates retrieval from judgment.

A full-filing evidence workflow where retrieval proposes candidate passages and a separate validator applies explicit risk thresholds.

ABroad signalHigher recall
BExplicit riskBalanced
CMaterial or activeMost selective

Insufficient evidence returns no result.

Key judgment Make the evidence standard, citation, and refusal behavior visible to the analyst.

Case study Public code
Architecture first: the project demonstrates a testable operating logic, not a claimed benchmark.
04Applied data science

Sovereign debt early warning

Testing whether alternative signals improve rare-event prediction.

I combined macroeconomic indicators, satellite nightlights, and news features using lag-only inputs and out-of-time validation.

Validation specificationAverage precision
Macro only0.461
Macro + satellite0.519

2019–2020 out-of-time split · offline validation

Key judgment Use average precision as the main comparison because crisis targets are rare.

Case study Public code
Measured result: model validation, not evidence of deployment or policy impact.

Operating approach

Strategy becomes useful when its decision boundaries are explicit.

01

Frame the workflow

Start with the user, the decision, and the constraint—not the model.

02

Build the critical path

Implement enough of the data, AI, and interface layers to test the operating idea.

03

Define the evidence bar

Separate measured results from estimates, and system behavior from substantive accuracy.

04

Design the handoff

Make uncertainty, review, escalation, and refusal part of the workflow.

Background

Business context, policy analysis, and technical implementation in the same conversation.

My experience spans international development research, strategy consulting, business analysis, and applied data science. I am most interested in the point where a technical capability has to become a credible way of working.

More about my background

Contact

Interested in AI implementation, strategy, or applied analytics?

Email LinkedIn GitHub Résumé