SkylerPathfinder
Transfer credit decisions your registrar can stand behind.
We built a PLAR and course articulation platform for colleges and universities. It reads syllabi, outlines, and resumes, classifies every learning outcome against Bloom's Taxonomy, and produces a side-by-side comparison report the registrar can sign off on. What took weeks of manual review now lands in seconds, with a five-metric statistical report attached to every decision.
A platform that gives registrars back their week, without giving up defensibility.
The brief
The client came to us with a problem every registrar recognizes. Transfer credit and Prior Learning Assessment decisions take weeks of manual syllabus comparison, judgments vary between evaluators, and when a student appeals, nobody has a written basis for the call. They wanted a platform that institutions would actually deploy, not a research tool. We built it, shipped it, and launched it publicly at skylerpathfinder.com.
Why most matching tools fail the registrar
A single fuzzy score is not enough. A registrar facing an appeal needs to show exactly where two courses align, where they diverge, and how confident the platform was at each level of Bloom's Taxonomy. Keyword matching cannot do that. Generic AI scoring cannot do that. What can: structured extraction from the source document, outcome-level classification, and a combined statistical report that treats each comparison as an ordinal-aware distribution, not a yes or no.
Who this is for
Registrars processing articulation requests. Deans trying to keep credit policy consistent across departments. Academic leaders who want to open more pathways for transfer and mature students without hiring a larger review team. The platform is multi-tenant, so an institution can manage its own outline library, run evaluations in-house, and share a scoped copy of any single evaluation with an applicant without exposing the rest of its data. Delivered across three stacked contracts, all 5.0 rated, POC through production through public launch with no gap.
POC through production through public launch, with no break between phases.
What we did
- A multi-tenant PLAR and course articulation platform live at skylerpathfinder.com
- Bloom's Taxonomy classifier with confidence scoring and ambiguity flagging
- Five-metric comparison engine that produces a defensible registrar-ready report
- Institutional outline library, course-articulation dashboard, and PLAR resume-to-course dashboard
- Three-tier access model: admin, institutional owner, and scoped student visitor
Our process
We started with the hardest piece. Could a platform classify a learning outcome against Bloom's Taxonomy with enough confidence to back a real credit decision? We built the classifier, the structured extractor, and the first version of the comparison engine, then tested it end to end against real syllabi.
With the core working, we hardened it into a product. Multi-tenant outline libraries, cached embeddings, OCR-aware document parsing, dashboards for articulation and PLAR, and controlled access for admins, evaluators, and students. This is where the platform absorbed the complexity that makes it institutional-grade.
We connected the platform to skylerpathfinder.com, added the scoped-visitor flow for sharing evaluations with applicants, wired up feedback capture, and shipped to production on Vercel.
Services covered
A stack built for source-of-truth ingestion, Bloom-aware scoring, and defensible comparisons.
- Next.js 15, React 19, TypeScript
- Dashboards for course articulation and PLAR
- Outline library and comparison workspace
- Scoped visitor view for shared evaluations
- Framer Motion for report walkthroughs
- Express 5 backend
- Schema-validated extraction endpoints
- Compare engine with custom concurrency limiter
- Three-tier JWT auth (admin, owner, visitor)
- Document upload handling with OCR diagnostics
- MongoDB and Mongoose
- Outline schema with performance caching built in
- Normalized topic sets, tokenized outcomes, precomputed embeddings
- Search across course names, topics, outcomes, and descriptions
- Content-level deduplication, so the same syllabus is never processed twice
- OpenAI API for extraction, embeddings, and comparison reasoning
- Pinecone for vector search
- Bloom's Taxonomy classifier (rule plus LLM)
- Ordinal-aware similarity stack with five combined metrics
- Confidence scoring and ambiguity flagging throughout
- PDF with pdf-parse and OCR diagnostics
- DOCX with mammoth
- Resumes, syllabi, and course outlines, same pipeline
- Parser diagnostics surfaced (char count, alpha ratio, OCR flag)
- Content-hashed so the same document never gets reprocessed
- Comparison engine that holds up under institutional load
- Performance caching so repeat comparisons skip the LLM entirely
- Wasserstein-1D, Spearman, JS-divergence, cosine, confusion matrix
- Side-by-side alignment, gap analysis, and source citations
- Every output is exportable as a registrar-ready report
Deployment pipeline
Stack summary
- Structured extraction firstLearning outcomes, topics, credit value, instructional hours, and textbooks are pulled into a validated schema, not kept as loose text.
- Bloom's classification per outcomeA 515-verb dictionary plus LLM margin scoring. Ambiguous outcomes get flagged for human review instead of silently misbucketing.
- Ordinal-aware similarityWe treat Bloom levels as an ordered scale, not categories. That is why a Level 4 outcome reads as closer to Level 3 than to Level 1, which matches how a registrar would score it.
- Five metrics combined, not oneWasserstein, Spearman, JS-divergence, cosine similarity, and a Bloom confusion matrix are all surfaced. The registrar sees how the platform reached its conclusion, not just the conclusion.
- Source-cited evidenceEvery alignment claim points back to the outcome text it came from, so a registrar can verify the call against the original syllabus in one click.
- Confidence, not yes or noEvery classification carries a confidence score. Borderline cases are flagged, not hidden, so the human reviewer knows exactly where to look.
- Three-tier accessInstitution admins manage the outline library, evaluators run comparisons, and students receive a scoped link to their own evaluation without ever seeing the rest of the library.
- Caching built in at the schema levelNormalized topic sets, tokenized outcomes, and mean embeddings are precomputed per outline. A repeat comparison skips the LLM entirely.
- Concurrency-aware OpenAI callsRate-aware batching with graceful retry and recovery keeps the comparison engine responsive under real institutional load.
- Institutional-scale searchSearch is tuned across course names, topics, outcomes, and descriptions, so the outline library stays responsive the moment it gets to ten thousand entries.
Key integrations
A launched platform. A registrar's workflow, rebuilt around defensibility.
A launched product for the client. Live at skylerpathfinder.com, in production use, and covered by three stacked 5.0-rated contracts from POC through public launch.
A registrar workflow that holds up on appeal. Every transfer credit or PLAR decision comes with a side-by-side comparison, source-cited evidence, and a five-metric statistical report. No more single fuzzy scores.
More articulation capacity, same team. Evaluators spend minutes on what used to cost them a week, so the institution can open more transfer pathways without growing the review team.
Institutional control, built in. Outline libraries, three-tier access, scoped visitor links, and dedupe per owner mean an institution can run evaluations in-house without leaking the library outside its walls.
- Course articulation dashboard for registrar-led transfer credit decisions
- PLAR dashboard that maps a resume straight to course equivalences
- Outline library with search, dedupe, and precomputed embeddings
- Scoped visitor flow for sharing a single evaluation with a student
- Side-by-side comparison report with alignment, gap analysis, and citations
- Five-metric statistical backing on every decision
- Confidence scoring so ambiguous outcomes are flagged, not hidden
- Support for PDF, OCR-suspect PDF, and DOCX in one ingestion path
Innovations
Most tools collapse course comparison to one fuzzy number. We combined Wasserstein-1D, Spearman rank correlation, Jensen-Shannon divergence, cosine similarity, and a Bloom confusion matrix. Each metric answers a different question, and together they produce a report a registrar can actually defend.
A 515-verb curated dictionary backs up the LLM. Every outcome classification carries a confidence score, and ambiguous ones are flagged rather than forced into a bucket. Human reviewers always know where to look.
Every claim in the comparison report points back to the outcome text it came from. A registrar can verify a call against the original syllabus in one click, which is what makes the decision defensible on appeal.
Normalized topic sets, tokenized outcomes, and mean embeddings are precomputed per outline. Repeat comparisons skip the LLM entirely, so running a thousand articulations stays affordable instead of blowing up the OpenAI bill.
- Quality is a release gate, not a promise. Every deployment carries parser diagnostics, cache hit rates, and classifier confidence into the data model, so the registrar never has to take the platform's word for it.
- Adding a new institution became a days-long exercise instead of a months-long one. The tenant model and the outline library absorb the variance.
- Every comparison has a paper trail. When a student appeals, the registrar has the five-metric report, the source citations, and the confidence scores in one place.
- Side-by-side learning outcomes with alignment scores
- Bloom-level confusion matrix across the pair
- Gap analysis, flagged outcomes, and coverage summary
- Source-cited evidence for every alignment claim
- Confidence score per classification
- A combined similarity readout across five metrics
- Exportable report the registrar can hand a student or a dean
- Its own outline library, managed in-house
- Three-tier access for admins, evaluators, and scoped visitors
- Per-owner dedupe so the same syllabus is never processed twice
- A feedback channel for evaluators back to the product team
- An ingestion path that accepts the documents they already have
- Scoped visitor links to share a single evaluation without exposing the library