Most resume parsing does not need AI.

Why it works this way

Every document that arrives takes the cheapest path that can answer it. If the exact text has been seen before, the structured result comes back from cache with no model call at all. If not, a local rule engine — regex, a learned skills gazetteer, and a growing set of section-heading synonyms — extracts what it can and scores its own confidence. Only when that confidence falls below threshold does the pipeline call out to Gemini, OpenAI or Anthropic.

The important part is the last step. Whatever the model returns is fed back into the local engine as new vocabulary: skills it had not seen, headings it did not recognise, formats it mis-read. The system that handles next month’s resumes is measurably better than the one that handled last month’s, and it calls the model less often to do it. Cost per document falls as volume rises, which is the opposite of how most AI tooling behaves.

What we are trying to build

A talent layer that is boring in the right places. Parsing should be fast, deterministic where it can be, and explainable when it is not. A match score should come with the reasons behind it — the skills present, the skills missing, the fit across experience, seniority, education and location — because a shortlist that can not be explained to a hiring manager is not worth much.

And when the system does generate text — a job description, a resume rewrite — it should be grounded. The Resume Builder checks every proposed rewrite against the candidate’s own resume before offering it, and blocks invented employers, skills and metrics rather than publishing them. That constraint costs us some fluency. We think it is the right trade.

Principles

• Local first. The model is a fallback, not the architecture.
• Grounded output. Generated text is checked against the source before a human ever sees it.
• Explainable scores. Every ranking comes with the evidence that produced it.
• Your data stays yours. Documents processed through TalentGraph are never used to train a model. See the security overview.
• Cheaper over time. If our cost per document is not falling, something is wrong with the design.

The team

TalentGraph is a small team working on talent infrastructure. If that sounds like something you want to work on, get in touch — we would rather hear from you than post a job ad.