Equity-centered college matching for the students who need it most.
🏆 First Place for Most Creative Project in the Educational Equity Track (sponsored by Snowflake), 7th Annual Datathon for Social Good @ UC Berkeley
Live demo · Quick start · How the matching works · Architecture
Top matches for a first-generation Latina student-parent in California with a $12k budget, public schools only: UC Davis ($8,594 net price for this family, 49% admission rate, 87% graduation rate), then CSU San Bernardino, a Hispanic-Serving Institution at $3,298 with a 91% admission rate. Ranked by what she'd actually pay and her odds of getting in and graduating, not prestige.
College search tools rank schools by prestige. But for a first-generation student, a student-parent, or a student from a low-income family, the question isn't "what's the highest-ranked school I can get into." It's "where will I actually graduate, without drowning in debt?"
The data says that question matters. In the federal datasets behind this project:
- Schools serving the highest-income students graduate 74.5% of them in 6 years; schools serving the lowest-income students graduate 34.7%, a 39.8-point gap.
- The median student-parent faces a $26,931 annual affordability gap once childcare is factored in, nearly 3× the $9,610 gap for other students.
- Black students graduate at rates ~12 points below white students at the same set of institutions.
EquiPath flips the ranking: instead of prestige, it scores every U.S. institution on how likely this specific student is to afford it, be supported there, and graduate.
A Streamlit app with five connected pages sharing one student profile:
| Page | What it does |
|---|---|
| My Profile | Structured form covering academics, budget, background, geography, and campus preferences |
| Chat Assistant | Builds the same profile conversationally, with optional voice input/output |
| My Recommendations | Ranks 4,933 institutions against your profile, each labeled with its K-means archetype; Claude explains each match in plain language, and a Q&A chat answers follow-ups |
| School Map | 6,812 schools on an interactive map, filterable by state or by your recommendations |
| Data Insights | The income-vs-graduation-rate analysis that motivated the project |
Every institution gets seven normalized scores, computed from federal data (no opinions, no rankings):
| Dimension | Default weight | Built from |
|---|---|---|
| Affordability | 25% | Net price + affordability gap (childcare-adjusted for student-parents) |
| ROI | 20% | Median earnings 10 years after entry vs. median debt of completers |
| Equity | 18% | Graduation rate for the student's own demographic group + cross-group parity |
| Support | 13% | Student-faculty ratio, instructional spend, endowment, Pell experience, retention |
| Academic fit | 13% | Program strength in the student's intended field, research intensity |
| Environment | 6% | Size/setting match + student-body diversity |
| Access | 5% | Admission likelihood given the student's GPA and test scores |
The weights adapt to the student: they're seeded from stated priorities and shift for circumstances (a student-parent's affordability score uses the childcare-adjusted gap; a first-gen student's support score amplifies high-support schools). Hard constraints (budget, in-state, institution type, minority-serving institution preference, distance from home zip code) filter before any scoring happens, and for-profit institutions are excluded by default.
A side effect worth noticing: highly selective schools sometimes surface for low-income students, not because of prestige but because full-need financial aid makes them among the cheapest options in the dataset (Stanford's net price for the lowest income bracket is effectively $0). The Access score keeps expectations honest by scoring admission odds near zero at such schools, and students who prefer a likelier path can filter or reweight accordingly.
Two design rules kept us honest:
- The algorithm decides, the AI explains. Rankings come from deterministic, auditable scoring; Claude is only used to translate the numbers into plain language and answer questions about the results.
- Demographics only ever help. Race/ethnicity is optional and used solely to surface the graduation rate for the student's own group and to match minority-serving institutions.
git clone https://github.com/alextoohey/EquiPath.git
cd EquiPath
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
streamlit run Home.pyThat's it: the datasets ship with the repo, and the first run builds a Parquet cache (~30 seconds; instant after that).
The AI features (match explanations, Q&A chat) are optional. To enable them:
cp .env.example .env # then add your Anthropic API keyRun the tests with pip install pytest && python -m pytest tests/.
The app runs as-is on Streamlit Community Cloud (free): point it at this repo with Home.py as the entrypoint, and add ANTHROPIC_API_KEY under app secrets if you want the AI features.
| Source | Contents |
|---|---|
| College Results View 2021 (IPEDS/College Scorecard) | Graduation rates by race, earnings, debt, admissions, institutional characteristics for 6,289 institutions |
| Affordability Gap AY2022-23 | Net price and affordability gaps per institution across five family-income brackets, including student-parent scenarios with childcare costs |
| NCES Postsecondary School Locations 2022-23 | Coordinates for the 6,812 schools on the map |
The two tabular datasets are joined on UNITID (the federal institution ID) and filtered to the income bracket matching the student's family income, yielding one row per institution: 4,933 institutions with ~290 features.
Home.py Streamlit entrypoint (landing page)
pages/ The five app pages
src/
├── data_loading.py Load + merge the datasets (Parquet-cached)
├── features.py All feature engineering → the 7 base indices
├── profile.py UserProfile dataclass + validation
├── scoring.py Hard filters + personalized scoring + ranking
├── clustering.py K-means institution archetypes
├── distance.py Zip-code radius filtering
├── llm.py Claude explanations for ranked matches
├── chat.py Conversational profile builder (+ voice)
├── profile_state.py Shared profile across pages (session state)
├── profile_editor.py Profile form UI
└── config.py Environment/API-key handling
tests/ Pipeline smoke tests
notebooks/ EDA that shaped the scoring design
flowchart LR
A[College Results 2021] --> C[data_loading<br/>UNITID join]
B[Affordability Gap<br/>AY2022-23] --> C
C --> D[features<br/>7 indices per school]
D --> E[scoring<br/>filters + personalized weights]
F[Student profile<br/>form or chat] --> E
E --> G[Ranked matches]
G --> H[Claude<br/>plain-language explanations]
G --> I[Map + visualizations]
- Persistent profiles: session state currently lives in memory, so a browser refresh starts over; a database plus lightweight auth would make profiles and saved searches durable.
- A learned success model: with outcome data (graduation, debt, earnings for students like this one at schools like this one), a supervised model could join the seven indices as an eighth, predictive score.
- More data: admission-rate coverage is ~35%; transfer pathways and campus support services (childcare, mental health) would sharpen the support index.
Built by Alex Toohey, Andre Lee, Davyn Paringkoan, and Leo Zhang for the 7th Annual Datathon for Social Good.
Data from the U.S. Department of Education (College Scorecard, IPEDS) and the Affordability Gap project. AI explanations powered by Anthropic's Claude.