Loading

Clearway Phoenix started after a simpler V1 matching calculator failed. The original system could only do literal keyword filtering, which meant a student looking for economy would miss programs labeled business or finance. That failure exposed the real problem: the agency was trying to make high-stakes placement decisions from messy spreadsheets, shallow logic, and manual cross-referencing.

I rebuilt the workflow as a sequence of connected systems: a normalized data model, a resumable AI-powered import pipeline, and a real-time matching engine that could be used during live advisor calls.

Rebuilding the data model

The first decision was sequencing. I did not start by importing data. I started by building the place that data needed to live.

The legacy system was too flat to support intelligent matching. Critical business rules lived inside notes fields: nationality restrictions, scholarship variants, language thresholds, fee structures, and accommodation details. That makes data display possible, but it makes reasoning impossible.

So the first pass rebuilt the foundation as a normalized schema across universities, programs, scholarships, locations, and reference data. The goal was not elegance for its own sake. The goal was queryability. If a future recommendation engine needed to reason about grade requirements, scholarship coverage, or field alignment, those things had to exist as real structured relationships first.

Decision

30-table normalized schema	Turned notes-field chaos into data the platform could actually query, filter, and score.
Major categories	Gave the future matching engine a structural relevance axis beyond raw keywords.
Typed pivots for requirements	Kept thresholds on the actual relationship where they belonged.
Embeddings stored on program rows	Avoided premature vector infrastructure while preserving semantic capability.
AI autofill in admin forms	Reduced dependency on manual operator discipline during ongoing maintenance.

ETL metric	Value
Universities	175
Accommodations	590
Programs	19,299
Scholarships	13,548
OpenAI calls on the completed run	16,363
Token volume	36.4M in / 9.3M out
Cache hits	21,073
Automated retries handled	640
Completed run cost	$11.05

Matching layer	Purpose
Hybrid search	Surface relevant candidates from 19,299 programs without exact-keyword dependence.
13-factor scoring model	Rank candidates with weighted business and eligibility logic.
Explainability panel	Show why a recommendation scored well or poorly.
18-status assignment lifecycle	Carry the workflow forward after discovery into actual placement operations.

Before	After
Advisors manually cross-referenced spreadsheets for hours per student	Advisors received ranked recommendations in under one second
Data lived in notes fields and inconsistent cells	Data landed in a structured schema built for filtering and scoring
Matching logic was shallow and literal	Matching combined semantic search with explainable weighted ranking
Recommendation quality was opaque	Every score exposed its reasoning, penalties, and strengths
Work stopped at discovery	Assigned programs entered a tracked lifecycle with prioritization and status changes

Rebuilding the data model

Building something similar?

Making the import resilient

From search to assignment

Results

Takeaway

Program Matching Engine

Rebuilding the data model

Building something similar?

Making the import resilient

From search to assignment

Results

Takeaway