Education Analytics Platform
Architected and deployed a production analytics platform (file-based Parquet storage; no database server) for a St. Louis public charter network. Sole developer. Engineered integrations with Canvas LMS, DeansList, NWEA MAP Growth, exit-ticket trackers, and a student-support roster into a unified data lakehouse — replaced manual state-compliance reporting and provided dashboards and reports to teachers, administrators, compliance directors, and the board.
Bilingual Family Progress Profile
A family-facing student profile the platform generates from the lakehouse: bilingual English/Spanish narrative, grading-cycle picker, NWEA MAP growth visuals, MTSS supports, and an accessibility-first design (WCAG-audited contrast and text sizing, dark/light themes). In any shared example, every name, score, and date is synthetic — no real student data appears.
Data sources
- Canvas LMS
- DeansList
- NWEA MAP
- Exit tickets
- Support roster
ETL pipeline
- Async fetchers
- Validators
- Transformers
- Schema checks
Lakehouse + ML
- Parquet lakehouse
- Star schema
- Fact + dim tables
- ML pipeline
- Growth predict
- Risk scoring
- Bias audits
REST API
- FastAPI
- 20 routers
- JWT auth
- RBAC middleware
Consumers
- Web dashboards
- 7 dashboards
- 12 KNet tabs
- Google Sheets
File-Based Lakehouse Architecture
Designed a file-based Parquet lakehouse that eliminated database dependencies. Runs on any machine with Python — no Postgres, no Redis, no Docker required.
Multi-Source ETL Pipeline
Async fetchers pull from Canvas REST API, DeansList API, NWEA MAP CSV exports, exit-ticket Google Sheets, and a student-support roster — replacing manual data collection with full incremental and snapshot modes.
Star-Schema Lakehouse
Fact and dimension tables organized in a star schema across 70+ courses. Supports time-travel via snapshot versioning with incremental and full-refresh modes. Sub-second analytical queries via DuckDB over Parquet.
ML Growth & Risk Pipeline
Deployed a 95%+ accurate gradient-boosted ensemble (XGBoost/LightGBM) trained on 14 risk indicators to predict growth trajectories and flag at-risk students. Model selection used stratified k-fold cross-validation to guard against label imbalance, with Optuna hyperparameter search and SMOTE-NC for minority-class oversampling. A 3-layer equity debiasing architecture (Fairlearn) runs post-hoc fairness audits and automated SHAP explanations at each inference cycle, monitoring for disparate impact across demographic cohorts before any intervention is assigned. Reduced crisis response time from 48 hours to 8 hours.
RBAC & JWT Auth
Role-based access control with JWT tokens. Teachers see their classes, admins see their campus, board sees network-wide. FERPA-compliant by design.
FastAPI REST Layer
Implemented 20 API routers serving dashboards, Google Sheets integrations, and admin tools. Auto-generated OpenAPI docs. Optimized queries to deliver sub-100ms response times on cached data.
7 Interactive Dashboards
Created teacher, admin, compliance, and board dashboards with drill-down filtering. Delivered real-time charts powered by Chart.js with responsive mobile layouts.
12 KNet Tab Views
Network-wide tabs surface cross-campus comparisons, attendance trends, behavior analytics, assessment performance, and MAP growth trajectories.
Google Sheets Sync
Bi-directional sync pushes live data to Google Sheets for stakeholders who prefer spreadsheets. Service-account auth with automatic token refresh.
Comprehensive Test Suite
Engineered a 350+-file pytest suite covering ETL transforms, API endpoints, ML pipeline outputs, and schema validation. Fixtures generate realistic synthetic data.
CI/CD Pipeline
28 GitHub Actions workflows for linting, testing, type-checking, schema validation, deploy previews, and production releases. Branch protection enforced.
Schema Validation & Contracts
Pydantic models enforce data contracts at every pipeline boundary. Schema drift detection alerts on upstream API changes before they break dashboards.