PROJECT — 2026

An AI college-admissions planning platform with 100+ active users — committee-grade application review without the private-consultant price tag.

Trajecta is a full-stack admissions planning platform I co-founded and lead engineering for, now serving 100+ active users: a React/Vite front end over an Express/TypeScript API on Render, with Supabase for Postgres and auth. It exists because committee-quality application feedback is mostly locked behind private consultants.

The core is a mock admissions committee. Four LLM officer readers — academic, narrative, impact, context — read the application in turn, each seeing the reports before it, and commit to a verdict on a five-point deny-to-admit scale. Fixed JSON schemas keep them arguing inside guardrails rather than free-associating. A chair reads only their reports, not the application, and decides admit, waitlist, or reject — and says where the room split. An eval harness replays real outcomes from public applicant self-reports: across 201 profile–school pairs, the committee called admit vs. non-admit correctly 82.6% of the time, against a 65% always-reject baseline. Each committee is briefed with the school's own Common Data Set factor ratings, and a separate chancing model sorts schools into reach, target, and safety.

Documents enter through an import pipeline that treats every upload as untrusted: PDFs are checked for type, size, and page count before a model sees them, the extraction is forced into a fixed JSON shape, and everything that comes back is sanitized and capped before it touches the profile. Nothing uploaded is stored.

Around it: GitHub Actions CI (lint, type-check, 78 Vitest test files), Stripe subscription billing, and rate limiting on every LLM call.

React · TypeScript · Express · Supabase · Anthropic API