AI implementation · Strategy · Applied analytics

Designing evidence-based AI and data systems for real decisions.

I am a Columbia SIPA MPA candidate with experience at the World Bank, Bain, and Capgemini. I work across problem definition, technical implementation, validation, and the operating choices required to make a system usable.

View case studies Download résumé

Selected work

Four cases, four kinds of evidence.

Real organizational workflow design, end-to-end analytical products, failure-aware AI architecture, and time-aware model validation.

01 · AI implementation

A reviewable AI workflow for cross-country legal research

For the World Bank’s Women, Business and the Law team, I designed and built a controlled pilot that separates source discovery, evidence checks, analysis, and human approval.

Contribution: Workflow architecture, agent logic, evidence routing, review controls, and validation
Evidence boundary: Completed pilot; operational impact and recurring use remain planned, not measured

Read case study

Figure 01Controlled evidence workflow

01Discovercandidate sources

02Checkscope and evidence

03Analyzereviewed pool only

04Approvehuman decision

Unverified evidence does not pass directly into a final answer.

02 · Decision system

Grid Intelligence: turning fragmented public data into an investigation workflow

I built a cloud-backed analytical product that connects U.S. electricity demand, generation, interchange, price, weather, and transition data across ten balancing authorities.

Contribution: Data acquisition, ETL, analytical logic, BigQuery layer, and interface design
Design choice: Expose interpretable signals and contradictions rather than opaque recommendations

Read case study

Figure 02Investigation view

Balancing authorityPJM · illustrative

Demand vs. forecast

Review

Interchange signal

Stable

Transition conditions

Mixed

Illustrative values; the case documents the implemented data and interface design.

03 · Applied AI design

A RAG system that separates retrieval from judgment

I built a prototype for identifying risk disclosures across full SEC 10-K filings. Retrieval proposes candidate evidence; a separate validation layer applies three explicit thresholds and may return no result.

Contribution: EDGAR ingestion, retrieval pipeline, threshold prompts, comparison logic, and interface
Design choice: Make evidence standards, source references, and refusal behavior visible to the analyst

Read case study

Figure 03Threshold comparison

ModeEvidence standardOutput

ABroad, forward-lookingHigher recall

BExplicit risk + contextBalanced

CMaterial or activeMost selective

If retrieved evidence does not meet a threshold, that mode returns an empty result.

04 · Applied data science

Testing alternative signals for sovereign debt early warning

I combined macroeconomic indicators, satellite nightlights, and global news signals, then used lag-only features and out-of-time validation to test what improved rare-event prediction.

Measured result: Macro + satellite average precision: 0.519 vs. 0.461 macro-only
Evidence boundary: Offline validation result, not evidence of policy adoption or causal impact

Read case study

Figure 04Validation average precision

Macro only

0.461

Macro + satellite

0.519

2019–2020 out-of-time validation · 258 country-years · 28 positive targets.

Working method

Four questions guide the work.

The goal is a credible decision process, not technical complexity for its own sake.

What is the decision?

Define the user, current constraint, and a useful outcome.

What must be built?

Translate the problem into a working flow, model, interface, or analysis.

What supports the claim?

Use relevant baselines, failure cases, and explicit evidence boundaries.

How would it be used?

Design handoffs, review points, uncertainty, and operating constraints.

Background

Business, policy, 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.

Discuss an opportunity or a problem worth implementing.

Email LinkedIn GitHub Download résumé