CareerCompiler · an AiGNITE portfolio build
A resume compiler with compile errors.
Your resume is compiled from a graph of career facts, each carrying its provenance and its span into the source document. Every rendered sentence cites the facts that license it. A sentence that cites nothing, carries a number found in no cited fact, or claims more than its evidence supports fails the build — scored against the cited facts by a pinned entailment model, with the score shown.
What it actually does
It gives honest verdicts. When a job posting carries a must-have requirement your fact graph cannot evidence, the verdict is do-not-apply, with the case against stated plainly.
Tailoring is evidence selection, not rephrasing. A solver chooses which facts make the page under a hard budget, and every omitted fact carries a typed reason. The language model that phrases the bullets never sees the job posting and restates facts faithfully — wording does not shift per job. That is a deliberate design decision, made after watching requirement-steered phrasing produce claims the evidence could not support.
The gate has no fallback. If the entailment model cannot run, the compile fails with a typed error. It never degrades to a keyword check, and no configuration can turn it off.
Published limits
On a golden suite of 4 labelled job cases against a 6-fact synthetic career graph, the analyzer places every requirement in the right bucket (matcher accuracy 1.0), returns the correct apply or do-not-apply verdict on every case including planted disqualifying gaps (1.0), never presents transferable evidence as direct (0 violations), and holds its verdict across pre-authored paraphrases of the job description (1.0); the suite is synthetic and small, so it does not measure performance on real resumes or real postings. The Phase 2 compiler rejects a sentence that claims more than its cited evidence supports: on the pinned entailment model a faithful sentence scored 0.9976 and an inflated verb 0.0015, but the 0.7 threshold is not yet calibrated between those extremes, so a borderline honest paraphrase can be rejected; compile-time rejection is measured, fabrication recall in the wild is not.
Not for
Not a cover-letter writer, not a keyword stuffer, not an interview coach. It will tell you not to apply.
API reference · source · the LLM senses, deterministic code decides, humans approve.