Inspiration

Every student asks some version of the same question: what should I be doing right now to set myself up for success? The usual answers are generic ("get an internship," "keep your GPA up") and rarely come with evidence. We had data on thousands of UMBC alumni and current students, including their majors, grades, internships, clubs, first jobs, and salaries. We wanted to use it to give real answers. We also didn't want to define "success" for anyone. A student aiming for the highest starting salary, one who wants to get hired quickly, and one who wants long-term growth need different advice.

What's your path? Find your success by following the Retrievers who came before you.

What it does

Success Metric Dashboard turns UMBC student and alumni data into guidance a student can act on. It has three parts:

An interactive story of the data. A single web page follows a student from their first day on campus to their career. Charts animate as you scroll, and each chart has hover tooltips, a table view, and a "Copy code" button that shows exactly how the chart was made. The page is split into three stages: Student Life (what students do outside class), Strategies (which choices in school pay off), and Job Trajectory (how graduates find their first job and what happens next).
Evidence-backed findings. Some of what we found:
- Major explains only about 2% of the variation in starting salary. The first job family explains about 20%, and the first employer's industry about 16%.
- The benefit of internships grows up to 2–3 internships and then levels off. Unemployment falls from 12.6% for alumni with no internships to 4.9% for those with three.
- A return offer from an internship adds roughly $5K–$7K to the median starting salary.
- Computer Science alumni start out earning slightly more than Information Systems alumni, but the IS alumni catch up by about year 5.
Advisor Ann. Ann is a personal advisor that students can talk to by typing, speaking, or uploading a document. She collects the student's profile, compares them with current peers and with alumni who became high earners, projects their likely job outcomes, and ranks the changes that would help most. Every recommendation comes from relationships that passed statistical significance tests.

How we built it

Analysis: Python with pandas, NumPy, SciPy, and statsmodels, done in Jupyter notebooks. We used ANOVA (η²) to see which factors explain salary, regression to find which factors hold up when considered together, and confidence intervals on proportions. We left out any group with fewer than 50 alumni so the medians would be reliable.
Data layer: We loaded all six datasets into Tiger Data (TimescaleDB on Tiger Cloud). Profile data like students, alumni, and courses stays in regular Postgres tables. Activity data like transcripts, experiences, and employment history is stored in hypertables with columnstore compression, and continuous aggregates on top make dashboard queries instant.
Advisor pipeline: A four-step pipeline: fit the student's profile to the dataset, measure their progress against peers, project their future using a regression plus their 50 most similar alumni, and rank strategy changes by expected salary gain.
Advisor Ann: Gemini turns what the student types, says, or uploads into structured fields, and ElevenLabs gives Ann her voice. Gemini only collects the profile. All of the actual advice comes from our tested pipeline, never from the language model. Ann's conversation memory and hourly usage stats are also stored in Tiger Data.
Frontend: A build script runs our analysis code and turns every chart into a JSON spec. A custom JavaScript renderer draws those specs as animated SVG charts, with 18 chart types including box plots, heatmaps, forest plots, and hexbins. The whole site builds into one HTML file with no dependencies, and every number on it comes from the same code as the notebooks. We also built a Streamlit dashboard for exploring the data.

Challenges we ran into

Defining "success." There's no single right metric, so we measured several outcomes (starting salary, salary growth, time to hire, unemployment) and let each student decide which matter to them.
Separating real effects from noise. Many relationships that look strong in a chart fell apart once other factors were taken into account. For example, GPA and major matter much less than they seem to. We decided the advisor would only recommend changes that held up under statistical testing.
Using an LLM without letting it make things up. A chatbot will confidently invent career advice. We limited Gemini to understanding what the student tells it, and left the recommendations to the statistics.
Small samples. Most alumni hold only a bachelor's degree, so comparisons involving master's degrees and some organizations had to be filtered or flagged.
Keeping everything in sync. Three notebooks, a Streamlit app, and a static site all show the same numbers. We had the site build its charts directly from the analysis code, so they can't fall out of sync.
Git. With several people rebasing branches and editing the same notebooks, we spent more time than we'd like to admit untangling merge conflicts in .ipynb files.

Accomplishments that we're proud of

Showing that your first job matters about ten times more than your major for starting salary, a finding that goes against common assumptions and that students can act on.
Building an advisor that can talk and listen, whose every recommendation is backed by a significance test.
A single-file site with 18 chart types we built ourselves, animated as you scroll, where every chart can show its data table and the code that made it.
An actual time-series database design in Tiger Data (hypertables, columnstore, continuous aggregates), not just a CSV dump.

What we learned

Interesting correlations are easy to find. The hard part is showing they're real and deciding what a student should actually do about them.
How to structure data by how it's used: fixed profile data and streams of activity events need different storage.
LLMs work best in our tool as the interface, not as the source of the answers.
Designing the charts carefully (what to show, what to leave out, how to explain sample sizes) mattered as much as the analysis itself.

What's next for Success Metric Dashboard

Live data: Connect to real, anonymized university records so the findings update every semester.
More outcomes: Grad school admission, job satisfaction, and long-term career growth, not just the first job.
Advisor integration: Give academic advisors a view of Ann's reports so they can review a student's plan together.
Other majors and schools: The pipeline is independent of any particular major, so it could extend beyond CS and IS, and to other universities.
Following students over time: Let students come back each semester and see how their progress and projected outcomes change.