Master Outline & Study Guide

Julius AI — U.S. Chronic Disease Indicators Analysis (AI-assisted EDA on CDC public-health data)

Purpose of this document: A complete, self-contained reference for studying this portfolio project. Use it to prepare for interviews — to explain what was built, how it works, why each decision was made, and to answer probing technical questions. When studying with Claude, this is the file Claude reads to teach and quiz you.

One sentence to anchor everything: This project uses Julius AI — a natural-language data-analysis tool — to explore the CDC’s 96 MB U.S. Chronic Disease Indicators dataset through six plain-English prompts, and the real skills on display are prompt engineering, public-health data literacy, interpretation, and critical evaluation of AI output — not manual coding.

Read §16 first if you’re prepping for an interview. An AI-assisted project invites a specific kind of skeptical question, and §16 is how you answer it well.

Table of Contents

1. The 30-Second Pitch
2. Why This Project Exists (Context)
3. What Julius AI Is (and What It Does Under the Hood)
4. The Dataset — CDC Chronic Disease Indicators
5. The Workflow & Prompt Engineering
6. Analysis 1 — Diabetes Prevalence: California vs. Texas
7. Analysis 2 — Asthma Prevalence by Race/Ethnicity
8. Analysis 3 — Physical Inactivity vs. Cardiovascular Disease Mortality
9. Analysis 4 — Mental Health Distress: Males vs. Females
10. Analysis 5 — Cigarette Smoking by Race/Ethnicity
11. Analysis 6 — National Obesity Trend
12. Key Findings & Recommendations
13. Julius AI — Tool Evaluation
14. Public Health Data Literacy (Concepts to Know)
15. Limitations & Honest Caveats
16. How to Position This Project in an Interview
17. Interview Q&A
18. How to Walk Through This Project Live
19. Glossary

1. The 30-Second Pitch

This project uses Julius AI — an AI-powered, chat-based data-analysis platform — to explore the CDC’s U.S. Chronic Disease Indicators (CDI) dataset: a 96 MB CSV with 800,000+ rows of chronic-disease surveillance data covering all 50 states, multiple years, and dozens of health topics.

The work was done entirely through natural-language prompts — no manual coding. Six analyses, each driven by one plain-English question: (1) diabetes prevalence trends in California vs. Texas, (2) asthma disparities by race/ethnicity, (3) the correlation between physical inactivity and cardiovascular-disease mortality, (4) the gender gap in frequent mental distress, (5) cigarette-smoking disparities by race/ethnicity, and (6) the national obesity trend. Julius generated the Python (pandas) code, ran it, and produced the charts; the project frames the questions, engineers the prompts, and interprets the output.

It is deliberately a two-part project: a public-health EDA and a structured evaluation of an AI analysis tool — when it works well, where it falls short, and when an analyst should and shouldn’t reach for it.

One-line version: “I used Julius AI to run six exploratory analyses on the CDC’s 96 MB Chronic Disease Indicators dataset through natural-language prompts — demonstrating prompt engineering, public-health data literacy, and a critical evaluation of where AI-assisted analysis is and isn’t appropriate.”

2. Why This Project Exists (Context)

The premise. Public-health agencies rely on chronic-disease surveillance data to find at-risk populations, allocate resources, and evaluate programs. But the CDC’s CDI dataset is large and technically demanding — 96 MB, hundreds of thousands of records, dozens of stratifications. Querying and visualizing it traditionally requires real SQL or Python skill. AI tools like Julius lower that barrier: a non-technical stakeholder can explore the same data through conversation.

So the project asks two questions at once:

1. A public-health question — what do six specific chronic-disease indicators actually show across states and demographics?
2. A tooling question — is an AI analysis platform a viable way to explore a large public-health dataset, and where are its limits?

Why it’s a legitimate portfolio project. This is the honest framing to internalize: the project does not claim to demonstrate coding ability. It demonstrates a different and genuinely valuable skill set — the things that remain hard even when an AI writes the code:

• Framing the right business question out of a messy 800k-row dataset.
• Prompt engineering — phrasing a question precisely enough that the AI produces a correct, targeted analysis.
• Public-health data literacy — knowing what age-adjusted prevalence means, what BRFSS is, why a stratification matters.
• Interpretation and data storytelling — turning numbers and charts into a coherent narrative with actionable recommendations.
• Critical evaluation — checking the AI’s output for statistical validity and knowing when not to trust it.

The project is current and relevant: it shows you can adopt and critically assess new tools, which is itself a sought-after analyst trait. (How to defend this in an interview is §16 — read it.)

3. What Julius AI Is (and What It Does Under the Hood)

Julius AI is a chat-based, AI-powered data-analysis platform. You upload a data file, ask questions in plain English, and it carries out the analysis and returns results and charts. The version used here, shown in the screenshots, is “Julius 1.0 Lite.”

What it actually does under the hood — know this; it’s the answer to “but you didn’t code it”:

Julius does not invent numbers. For each prompt it:

1. Generates Python code — the screenshots show a “Generated Code” / “Python” panel with real pandas code, e.g. df = pd.read_csv('U.S._Chronic_Disease_Indicators_20260206.csv', encoding='utf-8').
2. Executes that code in a sandboxed environment against the uploaded file.
3. Returns the result — tables, statistics, and matplotlib/plotly-style charts — plus a plain-language explanation.
4. Shows its work — the generated code is visible and inspectable, and a “Data Explorer” panel lets you browse the dataset.

So Julius is essentially an LLM that writes and runs pandas/Python data-analysis code on your behalf. It’s the same analysis a Python analyst would do — read_csv, filter, groupby, aggregate, correlate, plot — with the LLM translating English → code. That matters for interviews: the analysis is real, reproducible-in-principle Python; the AI is the interface, not a black box that fabricates results. The honest caveat (see §13, §15) is that you must read that generated code to confirm the method is sound.

Why a tool like this exists. It collapses the 30–60 minutes of manual coding a routine EDA task takes into a sub-minute conversational turn — which is genuinely useful for rapid exploration and hypothesis generation, the stage before formal analysis.

4. The Dataset — CDC Chronic Disease Indicators

What it is. The U.S. Chronic Disease Indicators (CDI), published by the Centers for Disease Control and Prevention (CDC). It’s a standardized, nationwide set of chronic-disease surveillance measures — the data public-health agencies use to track conditions like diabetes, asthma, cardiovascular disease, mental health, tobacco use, and obesity.

Scale. The file used was U.S._Chronic_Disease_Indicators_20260206.csv — 96 MB, 800,000+ rows, covering all 50 states plus territories, multiple years, and dozens of health topics.

The structure — one row per indicator-measurement. Each row is a single measured value for one indicator, in one place, in one year, for one demographic slice. The key columns:

Where the data comes from. Primarily the Behavioral Risk Factor Surveillance System (BRFSS) — CDC’s large telephone health survey — and the National Vital Statistics System (NVSS) — official death-record data (the source for mortality measures like cardiovascular-disease deaths).

The single most important structural fact — DataValueType. Because the same indicator can appear as a crude rate, an age-adjusted rate, a raw count, etc., every analysis must filter to the right DataValueType — almost always age-adjusted prevalence for fair cross-population comparison (see §14). Picking the wrong measure type silently produces a wrong answer. This is the dataset’s main “gotcha.”

The long/stacked shape. CDI is a “long” (tidy/stacked) dataset — one measurement per row rather than one column per metric. Analysis means filtering down to the Topic + Question + DataValueType + Stratification you want, then aggregating.

5. The Workflow & Prompt Engineering

The end-to-end workflow:

  CDC CDI dataset (96 MB CSV, 800k+ rows)
        │  uploaded directly to Julius AI
        ▼
  ┌─────────────────────────────────────────────────────┐
  │ JULIUS AI — one conversation per analysis            │
  │  1. Frame a business question                         │
  │  2. Write a natural-language prompt                    │
  │  3. Julius generates + runs Python (pandas) code      │
  │  4. Julius returns a chart + explanation               │
  │  5. (optional) refine with a follow-up prompt          │
  └───────────────────────────┬─────────────────────────┘
                              │  human interpretation
                              ▼
  ┌─────────────────────────────────────────────────────┐
  │ Insights → narrative → public-health recommendations  │
  │ + a structured evaluation of the tool itself          │
  └─────────────────────────────────────────────────────┘

Prompt engineering — the actual technical skill on display. The quality of the output depends heavily on the prompt. The six prompts show a deliberate progression from simple to highly specified. Compare:

• A looser prompt (Analysis 1): “How has the prevalence of ‘Diabetes’ changed in California vs. Texas over the last 5 years?” — names the topic, the locations, the time window.
• A highly specified prompt (Analysis 3): “Is there a correlation between states with high ‘Physical Inactivity’ and states with high ‘Cardiovascular Disease’ mortality? Make a scatter plot with a regression line with physical inactivity plotted on the x axis and cardiovascular disease plotted on the y axes and each point being a state.”

The lesson the project draws (and you should be able to state): the more the prompt specifies the method — the exact chart type, what each axis is, the unit of observation (“each point being a state”), the year — the more reliable and targeted the AI output. Vague prompts let the AI make assumptions; precise prompts constrain it. Good prompting is essentially writing a clear analysis spec in English.

Effective-prompt ingredients used across the six analyses: name the exact Topic/ Question in quotes, specify the locations or “entire United States,” pin the year(s), state the demographic stratification, and — when it matters — dictate the chart type and the axes. That precision is what made Julius reliably pick the right slice out of 800k rows.

6. Analysis 1 — Diabetes Prevalence: California vs. Texas

Business question. How has diabetes prevalence changed in California vs. Texas over the last ~5 years?

The prompt. “How has the prevalence of ‘Diabetes’ changed in California vs. Texas over the last 5 years?”

What Julius did. Filtered the CDI dataset to diabetes-prevalence records for CA and TX, pulled age-adjusted rates for 2019–2022, and produced a comparative line chart.

Key findings.

• Texas is consistently higher — ~11.2%–13.4% vs. California’s ~9.3%–10.9% across all four years.
• Both trend upward 2019→2022 (CA ~9.5%→10.7%; TX peaks at 13.4% in 2022).
• Texas 2021 dip then 2022 surge — a temporary decline to ~11.2% in 2021, then a jump to 13.4% — plausibly delayed reporting or post-pandemic diagnostic catch-up.
• A persistent 2–3 percentage-point gap — stable enough to suggest structural differences in demographics, risk factors, or healthcare access, not noise.

The teachable point. A consistent gap is more informative than a one-year difference — it points at a structural cause. Note also the honest hedge on the 2021→2022 spike (“potentially reflecting reporting effects”) rather than over-claiming.

7. Analysis 2 — Asthma Prevalence by Race/Ethnicity

Business question. Which racial/ethnic groups are disproportionately affected by asthma?

The prompt. “Break down ‘Asthma’ prevalence by Race/Ethnicity. Which groups are disproportionately affected?”

What Julius did. Filtered to asthma prevalence by racial/ethnic Stratification, aggregated median rates across states using the most recent available year per state, and produced a ranked summary.

Key findings.

• Multiracial, non-Hispanic highest — 15.6%, over 50% above the White, non-Hispanic rate of 10.1%.
• American Indian / Alaska Native, non-Hispanic — second highest at 12.6%.
• Black, non-Hispanic — 11.9%, above the overall average, consistent with established research on environmental exposure and healthcare-access disparities.
• Asian, non-Hispanic lowest — 1.0%, a large gap that may partly reflect data reporting differences, not only true prevalence.
• Hispanic — 8.3%, below the White rate, but this likely masks heterogeneity across Hispanic subpopulations.

The teachable point. Two pieces of analytical maturity here: (1) using a median across states to avoid one outlier state distorting a group’s rate, and (2) flagging that the strikingly low Asian figure and the aggregated Hispanic figure may be data artifacts — not accepting a surprising number at face value.

8. Analysis 3 — Physical Inactivity vs. Cardiovascular Disease Mortality

This is the most statistically substantive analysis — know its numbers cold.

Business question. Is there a correlation between states with high physical inactivity and states with high cardiovascular-disease mortality?

The prompt. “Is there a correlation between states with high ‘Physical Inactivity’ and states with high ‘Cardiovascular Disease’ mortality? Make a scatter plot with a regression line with physical inactivity plotted on the x axis and cardiovascular disease plotted on the y axes and each point being a state.” — the most specified prompt of the six.

What Julius did. Extracted per-state physical-inactivity rates and per-state cardiovascular-disease mortality rates, merged them by state, computed correlation statistics, and produced a scatter plot with a regression line annotated with the stats.

The statistics — memorize these:

• Pearson r = 0.721 — a strong positive correlation.
• R² = 0.52 — physical inactivity alone explains ~52% of the state-to-state variation in cardiovascular mortality.
• p < 0.001 — highly statistically significant; extremely unlikely to be chance.

Key findings.

• States above ~28% adult inactivity cluster at the highest cardiovascular mortality (220–260 deaths per 100,000).
• A clear gradient runs from ~16% inactivity / ~130 deaths per 100k to ~30% inactivity / ~240 deaths per 100k — a dose-response pattern.
• Policy implication: population-level interventions against sedentary behavior could meaningfully cut cardiovascular deaths.

The teachable point — and the one caveat you must volunteer: correlation is not causation. r = 0.721 is strong, but a cross-sectional state-level correlation can’t prove inactivity causes the deaths — confounders (age structure, income, healthcare access, diet) plausibly drive both. This is also an ecological analysis (states, not individuals), so an ecological fallacy risk applies: a state-level relationship doesn’t necessarily hold for individuals. Being able to say “strong, significant, but not causal, and ecological” is exactly the statistical literacy an interviewer is testing for.

9. Analysis 4 — Mental Health Distress: Males vs. Females

Business question. How do rates of frequent mental distress compare between men and women across the most populous states?

The prompt. “Compare the rates of ‘Frequent Mental Distress’ between men and women across the top 10 most populous states.”

What Julius did. Identified the top 10 states, filtered to frequent-mental-distress records stratified by gender for 2022, and produced a grouped bar chart with labeled percentages.

Key findings.

• Females higher in every state — gaps of ~4 to 9.5 percentage points; the female-exceeds-male pattern is universal across the states examined.
• Tennessee highest — 15.4% (male) and 24.9% (female) — nearly one in four women.
• Largest gap in Tennessee — 9.5 points.
• Regional pattern — Southern states (TN, AR, AL, MO) higher for both genders; NE, MN, MA lower.
• Nebraska lowest — 8.6% (M) / 15.7% (F) — but still shows the female-higher pattern, meaning the gender disparity holds regardless of baseline.

The teachable point — an honest data caveat the project itself flags. The prompt asked for the “top 10 most populous states,” but the screenshot notes Julius actually selected the top 10 states by data-record volume, which is a proxy for population, not true population rank. They overlap heavily but aren’t identical — a small example of the AI interpreting a prompt slightly differently than intended, and exactly why you must read what the AI did, not just the chart.

10. Analysis 5 — Cigarette Smoking by Race/Ethnicity

Business question. How does current cigarette-smoking prevalence vary across racial/ethnic groups?

The prompt. “Visualize the disparity in ‘Current Cigarette Smoking’ among different race/ethnicity categories for the year 2021.”

What Julius did. Filtered to current-cigarette-smoking records for 2021, stratified by race/ethnicity, pulled age-adjusted prevalence, and produced a horizontal bar chart ranked high to low.

Key findings.

• American Indian / Alaska Native, non-Hispanic highest — 27.1%, more than 2.6× the Hispanic rate.
• Multiracial, non-Hispanic — 19.3%, well above the national average.
• Black (15.7%) and White (15.0%) non-Hispanic — similar to each other.
• Hispanic lowest — 10.4%; the project notes this echoes the “Hispanic health paradox” in epidemiology (Hispanic populations sometimes show better-than-expected outcomes despite socioeconomic disadvantage).
• A ~17-point spread between highest and lowest → the case for culturally tailored cessation programs over one-size-fits-all.

The teachable point. Naming the “Hispanic health paradox” shows domain knowledge — it demonstrates you’re interpreting the number through public-health literature, not just reading a bar height.

11. Analysis 6 — National Obesity Trend

Business question. What is the national trend in adult obesity prevalence over time?

The prompt. “Plot the trend of ‘Obesity among adults’ from the earliest available year to the latest for the entire United States.”

What Julius did. Filtered to national-level (“United States”) adult-obesity records, aggregated by year, and produced a time-series line chart for 2019–2022.

Key findings.

• 2021 peak — ~34.9%, up from ~33.5% in 2020 — likely pandemic-era lifestyle disruption (less activity, more sedentary time, dietary change).
• 2022 partial reversal — back to ~33.5% as restrictions eased.
• Narrow range, alarming baseline — the whole span is only ~1.4 points (33.5–34.9%), but roughly one in three U.S. adults was obese throughout.
• No sustained improvement — 2022 (~33.5%) is essentially unchanged from 2019 (~33.6%); existing interventions have not bent the curve.

The teachable point. The headline isn’t the small year-to-year wobble — it’s that the baseline is ~33% and flat. A good analyst reads the level, not just the slope, and resists over-interpreting a 1.4-point range as a “trend.”

(Minor filename note for a walkthrough: the Analysis 6 image files are named Chart_6_Diabetes_Across_Nation.png / Question_6_… even though Analysis 6 is the obesity trend — a leftover naming slip, not a content error.)

12. Key Findings & Recommendations

The five public-health findings (the synthesis across all six analyses):

1. Chronic-disease burden varies sharply by state — Texas diabetes (~13.4%) ran 2–3 points above California (~10.7%).
2. Racial/ethnic disparities are pervasive — American Indian/Alaska Native and Multiracial populations carried disproportionate burden in both asthma and tobacco use.
3. Physical inactivity is a powerful predictor — r = 0.721 with cardiovascular mortality; population-level activity promotion could meaningfully cut heart-disease deaths.
4. Mental distress disproportionately affects women — females higher in every state examined, widest in the South.
5. Obesity is stuck high — ~33–35% nationally with no real downward trend.

Plus the sixth, meta-finding: AI-assisted analysis is viable for rapid exploration of a large public-health dataset.

The recommendations (what an agency should do): prioritize high-burden states (TX for diabetes, TN for mental health) for funding; build culturally tailored programs for the hardest-hit groups (especially AI/AN communities); invest in physical-activity infrastructure given the inactivity↔CVD link; expand gender-responsive mental-health services; and use AI tools for rapid initial surveillance and hypothesis generation ahead of formal statistical work.

The framing to remember: every analysis ends in a decision, not just a chart. The recommendations are appropriately scoped as directional — “prioritize,” “invest,” “explore” — because exploratory analysis supports hypotheses, it doesn’t prove causal policy effects.

13. Julius AI — Tool Evaluation

The project’s second half is a structured evaluation of the tool itself — and this is where the “critical thinking” skill shows. Be ready to discuss both columns.

What worked well:

• Handled the large dataset — loaded and queried a 96 MB / 800k-row CSV with no performance issues.
• Speed — each of the six analyses completed in under a minute through one conversational prompt; the equivalent manual Python work is ~30–60 minutes each.
• Automatic, well-formatted visualization — correct scales, labels, legends, and even on-chart statistical annotations, with no manual styling.
• Statistical awareness — correctly computed Pearson r, R², and p-values when asked, and applied sensible transformations (age-adjusted rates, median aggregation).
• Iterative refinement — follow-up prompts adjusted charts and drilled into subsets without rewriting code.

Limitations observed:

• Black-box methodology — Julius shows its generated code, but the user must read it to confirm the statistical approach fits the data and question.
• Prompt sensitivity — output quality varied with phrasing; specific prompts (axes, chart type, year) gave more reliable results.
• Reproducibility — exact outputs can vary between sessions, so scripted workflows are preferable for production or regulatory work.
• Interpretation still required — Julius produces numbers and charts; domain expertise is still needed to contextualize them.

When to use Julius (the project’s verdict):

The teachable point. The honest verdict — excellent for exploration, not for production or regulated work — is the most important single output of the project. It shows you can adopt a new tool and judge its proper scope, which is more valuable than uncritical enthusiasm.

14. Public Health Data Literacy (Concepts to Know)

An interviewer for an analyst role — especially in health, government, or insurance — may probe whether you understand the data, not just chart it. Know these.

Age-adjusted rate. A rate statistically standardized to a reference age distribution so two populations can be compared fairly. Florida (older population) and Utah (younger) will look different on a crude disease rate purely because of age; the age-adjusted rate removes that distortion. The CDI dataset offers this via DataValueType — and this project correctly uses age-adjusted prevalence for its comparisons. If asked “why age-adjusted?” — so the comparison reflects disease burden, not just demographic age structure.

Prevalence vs. incidence. Prevalence = the share of a population that has a condition at a point in time (what this project measures). Incidence = the rate of new cases over a period. Chronic-disease surveillance is mostly prevalence.

Crude rate vs. age-adjusted rate. A crude rate is the raw count over the population, unadjusted. Age-adjusted is the comparison-safe version. The DataValueType column distinguishes them — picking the wrong one is the dataset’s main trap.

BRFSS (Behavioral Risk Factor Surveillance System). CDC’s large annual telephone health survey of U.S. adults — the source for self-reported indicators like diabetes, asthma, smoking, inactivity, mental distress. Because it’s self-reported and survey-based, it carries sampling error and reporting bias.

NVSS (National Vital Statistics System). Official birth/death-record data — the source for mortality indicators like cardiovascular-disease deaths. Death records are far more complete and reliable than survey self-report.

Mortality rate “per 100,000.” Deaths are reported per 100,000 population so states of different sizes are comparable — the CVD mortality figures in Analysis 3 (~130–260 per 100k) are in these units.

Stratification. Splitting a measure by a demographic dimension — race/ethnicity, gender, or “Overall.” Disparities analysis is comparing across stratifications.

Ecological analysis / ecological fallacy. Analysis 3 correlates state-level aggregates. An ecological analysis uses groups, not individuals, as the unit. The ecological fallacy is wrongly assuming a group-level relationship holds at the individual level. A state-level inactivity↔mortality correlation does not prove an inactive person will die of heart disease.

Correlation ≠ causation. The single most important caveat for Analysis 3 — covered in §8. Strong, significant, but cross-sectional and confoundable.

15. Limitations & Honest Caveats

Volunteer these — they show analytical maturity and pre-empt the skeptical questions.

1. It’s exploratory analysis, not causal proof. Every finding is a pattern or association. Analysis 3’s r = 0.721 is a strong correlation, not evidence that inactivity causes cardiovascular deaths.
2. Ecological / aggregate data. State- and group-level analysis — subject to the ecological fallacy (§14); conclusions don’t transfer to individuals.
3. Self-reported survey data. Most indicators come from BRFSS, a telephone survey — so they carry self-report bias, sampling error, and coverage gaps. Surprising values (e.g., the 1.0% Asian asthma rate) may be data artifacts, and the project says so.
4. AI black-box risk. Julius generated the code; the analyst must read it to confirm the method. The project flags this explicitly as a limitation.
5. Reproducibility. AI outputs can vary between sessions — unsuitable for production or regulatory reporting. The project states this.
6. Prompt-interpretation drift. Analysis 4 is a concrete example — “top 10 most populous states” was answered as “top 10 by data-record volume,” a proxy. The AI’s interpretation must be checked against intent.
7. No raw data or notebook committed. The 96 MB CSV was uploaded to Julius, not stored in the repo; the project’s artifacts are the charts, prompts, and write-up. The analysis is not re-runnable from this folder alone.
8. Small time windows. Several analyses cover only 2019–2022 — short for genuine “trend” claims; year-to-year movement can be reporting noise as much as real change.
9. The skill shown is not coding. Stated plainly because it’s the honest framing — this project demonstrates prompt engineering, data literacy, interpretation, and tool evaluation. Coding ability is demonstrated by the other portfolio projects.

16. How to Position This Project in an Interview

Read this section carefully — an AI-assisted project invites one specific skeptical question, and how you answer it determines whether this project helps or hurts you.

The question you will get, in some form: “You didn’t actually write any code here — what does this project show me?”

The wrong answer: getting defensive, or overclaiming (“it’s basically the same as coding it myself”).

The right answer — three moves:

1. Name the skill honestly and reframe it as valuable. “You’re right that I didn’t hand-write the pandas code — Julius generated it. What this project demonstrates is the layer around the code: framing the right question out of an 800,000-row dataset, engineering prompts precise enough to get a correct analysis, reading the generated code to confirm the method, and interpreting the results with public-health context. Those are the parts that stay hard even when an AI writes the code.”

2. Prove you understand what’s under the hood. “Julius isn’t a black box that invents numbers — it writes and runs real Python and pandas. I can read that generated code, and I’d write the same read_csv → filter → groupby → correlate → plot myself — my SQL and Python projects show that. Here the point was the tool, not re-proving I can code.”

3. Lead with the critical evaluation. “Half the project is a structured evaluation of the tool — where it’s excellent (rapid EDA on big files) and where it’s not acceptable (production pipelines, regulated or reproducible analysis). I caught the AI quietly reinterpreting one of my prompts. Knowing when not to trust a tool is the skill I’d most want you to take from this.”

Why this project genuinely strengthens a portfolio (say a version of this):

• It’s current — fluency with AI analysis tools is now an explicitly sought trait.
• It shows range — paired with the SQL/Python/Excel projects, it proves you pick the right tool for the job rather than reflexively using one.
• It shows judgment — the willingness to write down a tool’s limitations is rarer and more valuable than tool enthusiasm.
• It shows domain literacy — age-adjusted rates, BRFSS vs. NVSS, the ecological fallacy, the Hispanic health paradox — that’s public-health knowledge, not coding.

The one-sentence positioning: “It’s an AI-assisted exploratory analysis and a critical evaluation of the tool — it demonstrates prompt engineering, public-health data literacy, and the judgment to know where AI analysis belongs and where it doesn’t.”

17. Interview Q&A

Practice these out loud. Answers are written the way you’d actually speak.

Q1. Walk me through this project. “I used Julius AI — an AI data-analysis platform — to explore the CDC’s U.S. Chronic Disease Indicators dataset, a 96 MB file with over 800,000 rows. I ran six analyses, each from a single plain-English prompt: diabetes trends in California versus Texas, asthma disparities by race, the link between physical inactivity and cardiovascular mortality, the gender gap in mental distress, smoking disparities by race, and the national obesity trend. Julius generated and ran the Python code; I framed the questions, engineered the prompts, and interpreted the results. The second half of the project is a structured evaluation of the tool itself.”

Q2. You didn’t write the code — what does this project actually demonstrate? “Fair question, and I’d answer it directly. It demonstrates the work that stays hard even when an AI writes the code: framing a good question out of 800,000 rows, prompt engineering precise enough to get a correct analysis, reading the generated code to verify the method, and interpreting results with public-health context. My SQL and Python projects show I can write the code; this one shows I can use an AI tool effectively and — just as important — evaluate where it shouldn’t be used.”

Q3. What is Julius AI actually doing under the hood? “It’s a large language model that writes and runs Python and pandas code for you. When I ask a question, it generates real code — I saw pd.read_csv on the CDC file, then filtering, grouping, aggregation, plotting — runs it in a sandbox, and returns the chart plus an explanation. The generated code is visible, so it’s not a black box that fabricates numbers. It’s the same analysis a Python analyst would do; the AI is the interface that translates English into code.”

Q4. How did you make sure the AI’s analysis was correct? “A few ways. I read the generated code to confirm it filtered to the right indicator and the right DataValueType — age-adjusted prevalence, which is the dataset’s main trap. I sanity-checked the numbers against what’s plausible in public-health literature. And I caught at least one prompt-interpretation drift — in the mental-health analysis I asked for the top 10 most populous states and Julius used the top 10 by data-record volume as a proxy. That’s exactly why you read what the AI did, not just the chart.”

Q5. What’s age-adjusted prevalence and why does it matter here? “It’s a rate standardized to a common age distribution so two populations can be compared fairly. Florida has an older population than Utah, so a crude disease rate would look worse in Florida purely because of age. Age-adjusting removes that, so the comparison reflects actual disease burden. The CDI dataset has a DataValueType column, and I made sure every comparison used the age-adjusted version.”

Q6. Tell me about the strongest statistical finding. “The physical-inactivity versus cardiovascular-mortality analysis. Across states, the Pearson correlation was 0.721 — strong and positive — with an R² of 0.52 and a p-value below 0.001. So inactivity alone explains about half the state-to-state variation in heart- disease mortality, and it’s highly significant. But I’d immediately add the caveat: correlation isn’t causation, it’s a cross-sectional state-level analysis, so confounders like age and income plausibly drive both — and it’s ecological, meaning the state-level relationship doesn’t necessarily hold for individuals.”

Q7. What’s the ecological fallacy, and does it apply here? “It’s the mistake of assuming a relationship that holds for groups also holds for individuals. Analysis 3 correlates state-level rates — states are the unit of observation. The ecological fallacy would be concluding that an inactive person will die of heart disease. The state-level link is real and useful for policy targeting, but I wouldn’t extend it to individual risk.”

Q8. What were Julius AI’s limitations? “Four main ones. It’s a partial black box — it shows the code but you have to read it to trust the method. It’s prompt-sensitive — vague prompts give worse results. Outputs aren’t perfectly reproducible between sessions, so it’s wrong for production or regulated work. And it produces numbers, not understanding — you still need domain expertise to interpret them. My verdict was: excellent for rapid exploration and hypothesis generation, not appropriate for production pipelines or regulatory reporting.”

Q9. When should an analyst use a tool like Julius, and when not? “Use it for the exploratory stage — profiling an unfamiliar dataset, quick hypothesis testing, fast presentation charts. That’s where the sub-minute turnaround is a real advantage. Don’t use it for production data pipelines or anything that has to be reproducible and auditable — regulatory reporting, for instance — because outputs can vary between runs. The right move is AI for exploration, then scripted Python or SQL for the formal, reproducible analysis.”

Q10. What was the most important public-health finding? “The pervasiveness of racial and ethnic disparities. American Indian and Alaska Native populations had the highest cigarette-smoking rate at 27.1% — more than two and a half times the Hispanic rate — and were also near the top for asthma. That kind of consistent, cross-condition disparity is the strongest argument for culturally tailored interventions rather than one-size-fits-all programs.”

Q11. How did prompt engineering affect your results? “A lot. The more I specified, the better the output. My loosest prompt just named the topic and locations. My most specified one — the inactivity-versus-mortality one — named both indicators, asked for a scatter plot with a regression line, and said each point should be a state and which variable went on which axis. That precision is essentially writing an analysis spec in English, and it produced a much more reliable result. Vague prompts let the AI make assumptions; precise prompts constrain it.”

Q12. What would you do differently or next? “For anything beyond exploration I’d take Julius’s generated code, move it into a proper Python notebook, and make it reproducible — pin the data version, set the filters explicitly, and re-run it as a script. I’d also want longer time windows; several of my trends only covered 2019 to 2022, which is short. And I’d treat these findings as hypotheses — the inactivity-mortality link, for instance, would justify a proper multivariable model that controls for confounders.”

18. How to Walk Through This Project Live

If asked to screen-share or talk through the project page, use this order:

1. State the dual purpose up front — “this is two projects in one: an exploratory public-health analysis, and an evaluation of an AI analysis tool.” Saying that first pre-empts the “you didn’t code it” question.
2. Show the dataset — the CDC CDI: 96 MB, 800k+ rows, the long/stacked structure, and the key columns. Emphasize DataValueType and why age-adjusted prevalence matters.
3. Show one Julius prompt screenshot — point at the generated Python code panel. Make the point explicitly: “it writes and runs real pandas; it’s not fabricating numbers.”
4. Walk the strongest analysis — Analysis 3 (inactivity vs. CVD mortality). Give the stats (r = 0.721, R² = 0.52, p < 0.001) and immediately give the caveat (correlation ≠ causation, ecological). Leading with the caveat is what signals statistical maturity.
5. Show one disparities analysis — asthma or smoking by race — and name a piece of domain knowledge (the Hispanic health paradox, or flagging the suspicious 1.0% figure as a possible data artifact).
6. Spend real time on the tool evaluation — the “what worked / limitations / when to use” verdict. This is the critical-thinking payload.
7. Close on the positioning — “AI for exploration, scripted code for production; the skill here is prompt engineering, data literacy, and knowing the tool’s limits.”

Pacing tip: don’t oversell the six analyses as if they were the hard part — they’re solid EDA. Spend your time on (a) what Julius does under the hood, (b) Analysis 3’s stats with its caveats, and (c) the tool evaluation. Those three show judgment.

19. Glossary

• Julius AI — a chat-based, AI-powered data-analysis platform; generates and runs Python/pandas code from natural-language prompts. Version used: “Julius 1.0 Lite.”
• Prompt engineering — crafting natural-language instructions precise enough to get a correct, targeted analysis from an AI tool.
• CDI (U.S. Chronic Disease Indicators) — the CDC’s standardized nationwide set of chronic-disease surveillance measures; the dataset analyzed (96 MB, 800k+ rows).
• CDC — Centers for Disease Control and Prevention; publisher of the dataset.
• BRFSS (Behavioral Risk Factor Surveillance System) — CDC’s large annual telephone health survey; the source of most self-reported indicators in CDI.
• NVSS (National Vital Statistics System) — official U.S. birth/death-record data; the source of mortality indicators.
• Prevalence — the share of a population that has a condition at a point in time (what this project measures).
• Incidence — the rate of new cases over a period (distinct from prevalence).
• Age-adjusted rate — a rate standardized to a reference age distribution so populations with different age structures can be compared fairly.
• Crude rate — an unadjusted raw rate (count ÷ population).
• DataValueType — the CDI column specifying which kind of measure a row holds (age-adjusted prevalence, crude rate, number, …); filtering to the right one is essential.
• Stratification — a demographic breakdown of a measure (race/ethnicity, gender, “Overall”).
• Pearson r — a correlation coefficient from −1 to +1; 0.721 here = strong positive.
• R² — the share of variance in one variable explained by the other; 0.52 = ~52%.
• p-value — the probability the observed relationship arose by chance; p < 0.001 = highly significant.
• Regression line — the best-fit line through a scatter plot summarizing the trend.
• Correlation ≠ causation — a relationship between two variables doesn’t prove one causes the other.
• Ecological analysis — analysis using groups (e.g., states) as the unit of observation rather than individuals.
• Ecological fallacy — wrongly assuming a group-level relationship holds for individuals.
• Confounder — a third variable that influences both variables in a correlation, potentially explaining the apparent link.
• Hispanic health paradox — the epidemiological observation that Hispanic populations sometimes show better-than-expected health outcomes despite socioeconomic disadvantage.
• EDA (Exploratory Data Analysis) — the initial, open-ended stage of investigating a dataset to find patterns and generate hypotheses.

This study guide documents the project as presented. The authoritative reference is the project page index.md and the prompt/chart screenshots in images/. The underlying 96 MB CDC dataset was uploaded to Julius AI and is not stored in this folder. When this guide and the project page disagree, the project page wins.