Neyda Larson | Data Analyst & AI Systems Builder

Data Analyst & AI Systems Builder · Case Studies

All the builds.

Las Vegas, NV · Remote

LinkedIn

GitHub

English · Spanish

Every data build and AI system I've shipped, in one place.

See how they work
Let's talk

DATA BUILDS

Real data · 2024

Full technical

The Data Quality Question a Federal Survey Left Unanswered

Pre-registered RCT analysis: survey-weighted jackknife estimation, Holm-corrected multiplicity testing, and a power analysis paired with every null.
  • Jackknife SE derived from NCI's spec, validated against a hand-computed fixture
  • An implied_n guard halts the pipeline on a statistically impossible estimate
  • Power (MDE) analysis paired with every null, not just a p-value

Python (pandas, numpy), R (survey package) for an independent cross-check, Power BI for the scorecard.

View case
View code

Disaggregates the aggregate

Analysis

Diagnosing a Segment Failure Hidden by Healthy Aggregates

Python and pandas: cohort splits, z-tests, and a six-tier validation framework isolate execution from market as the real cause.
  • Confidence intervals and z-tests on every headline comparison
  • Open-date vs. close-date cohort split reveals a reporting lag
  • Six-tier validation framework, including anti-outcome-encoding checks

Python, pandas, NumPy, Jupyter.

View case
View code

Real data · 2024

Full technical

Rural Hospital Risk Hidden by a Healthy Regional Average

Star schema in DuckDB and Power BI: DAX filter-context fixes and independent Python validation catch what regional averages hide.
  • Six-dimension star schema built from a 143,613-row extract
  • DAX filter-context bugs fixed with ALL(), KEEPFILTERS, CROSSFILTER
  • Every KPI re-derived independently in Python for validation

DuckDB · SQL · Power BI · DAX · Python for independent validation.

View case
View code

ai BUILDS

Builds memory

Tested & benchmarked

Operator Brain Evaluation

Meta-skill benchmark: a staged interview writes real files from real answers, tested against predictions locked before any run.
  • Read-before-write pattern guards against concurrent edits
  • Explicit scope fence checked before every decline
  • Predictions for 9 cases written before the skill was built

Claude · prompt and system design · eval harness

View case
View code

Helps you reflect

Tested & red-teamed

Decision Clarity

Six-stage conversation architecture red-teamed against real safety cases, with hard constraints enforced through prompt design.
  • Red-teamed against self-harm and escalation test cases
  • One-question-per-turn constraint enforced by rule, not luck
  • Benchmarked across 5 fixtures on pass rate, time, tokens

Claude · prompt and conversation design

View case
View code

Sorts leads instantly

Shipped · logic

Make.com Lead Classifier

No-code waterfall logic in Make.com: six boolean variables and a five-module pipeline classify and route leads safely.
  • Store-first design saves data before any routing logic
  • Boolean-to-text type coercion fixed at the mapping layer
  • Full test matrix, one crafted submission per classification path

Tally (form), Make.com (logic and routing), MailerLite (storage and segmentation).

View case
View code

Talks with people

Designed

Lead Magnet AI

Built with Claude API prompt engineering and a 7-stage conversational architecture, routed through Make.com into MailerLite and WordPress.
  • Prompt engineered for generated output, never templated
  • 7-stage conversation flow with built-in restraint rules
  • Make.com backend routes leads into segmented email flows

Claude API · Make.com · MailerLite · WordPress

View case
View code

© 2026 Redirect Coaching, LLC

Let's Talk