Four frontier LLMs and three ATS parsers read your resume in parallel. Tyr surfaces where they agree, where they disagree, and what that gap tells you about how you will be read.
Before a human opens your file, automated systems parse, score, and filter. Most candidates never learn what was extracted — or what was missed.
GPT-4o, Claude, Gemini, and Llama each summarize your experience, read your seniority, and flag gaps — all before a hiring manager sees your name. tyr measures where they agree and where they do not.
One upload. Two complete reports. Your parser-disagreement score, your AI-legibility score, the per-query σ and ρ across models, and the exact edits that move both numbers.
Before a human opens your file, automated systems parse, score, and filter. Most candidates never learn what was extracted — or what was missed.
GPT-4o, Claude, Gemini, and Llama each summarize your experience, read your seniority, and flag gaps — all before a hiring manager sees your name. tyr measures where they agree and where they do not.
One upload. Two complete reports. Your parser-disagreement score, your AI-legibility score, the per-query σ and ρ across models, and the exact edits that move both numbers.
Three steps
Drop your PDF. tyr accepts any layout — multi-column, tables, exotic fonts.
Three parsers and four LLMs run simultaneously, exactly like the hiring stacks you will apply to.
Disagreement scored, σ and ρ surfaced, AI-legibility quantified, and concrete edits flagged for both reports.
What you get
The structural reality and the interpretive read — side by side.
Sample findings
"Candidate presents as technically capable but lacks evidence of cross-functional leadership or measurable impact."
"Contact block failed to parse — email not extracted. Candidate likely filtered before scoring."
"Seniority read: mid-level. Applicant used no quantified outcomes in 6 of 7 bullet points."
Four heterogeneous frontier LLMs — GPT-4o, Claude Sonnet, Gemini, and Llama 3.1 70B. Each receives the same eight structured queries about your resume. We measure numerical disagreement (σ across scalar judgments) and reasoning dispersion (ρ across embedded explanations). Disagreement is treated as a calibrated uncertainty signal, not noise.
Three independent parsers run in parallel — Affinda's commercial NER, the open-source OpenResume engine, and our own deterministic extractor. We normalize their outputs into a canonical schema and score where they diverge. High parser disagreement is itself a finding: it predicts that real-world ATSes will read your resume inconsistently.
Two readings. High σ on a scalar query (e.g. seniority) means LLM-powered screeners will reach different conclusions about you depending on which one they use — your resume reads ambiguously. High ρ on reasoning text means the models are looking at different signals to arrive at their answer — your resume is multi-interpretable. Both are addressable with concrete edits.
Yes — encrypted at rest in a row-level-security-isolated Postgres instance keyed to your account. You can delete it at any time. We never train models on user data and never share your resume with third parties.
The disagreement score is robust by construction — if three parsers extract the same field, real ATSes overwhelmingly will too. The σ and ρ metrics are calibrated against a 5,000-resume reference distribution. We do not claim to predict any specific employer’s hiring decision; we measure how the AI layer of the funnel reads you, with explicit uncertainty.