Resume analysis
Reading documents as documents rather than pattern-matching against a keyword list.
01 / the problem
The documents arrive.The reading does not scale.
- Every resume has a different structure.
- Keyword filters reject people who can do the job.
- The first applications get read more carefully than the last.
- Nobody can explain afterwards why a candidate was ranked where they were.
Screening is a reading problem disguised as a volume problem. The information needed to make the decision is in the documents, in prose, in whatever format each applicant chose.
Keyword matching is fast and wrong. Reading everything properly is correct and does not scale.
Language models changed which of those two constraints is real.
02 / the decision
A product, not a chat window.
A general-purpose assistant can analyse one resume well. That is a demo, not a system.
The actual job is comparative and repeatable: score many candidates against one role, apply the same criteria to each, and produce output a hiring manager can defend.
That is a product shape, not a prompt.
Bulk ranking
The same criteria applied consistently across the whole applicant pool.
Job-fit evaluation
Candidates assessed against the requirements of a specific role, not a generic score.
Document Q&A
Asking direct questions of a document set instead of re-reading it.
03 / the product
Ranked, with the reasoning attached.
04 / the build
Where AI products usually break.
The model is rarely the hard part. Everything around it is.
01
Consistency across a batch
A ranking is only useful if the hundredth candidate is assessed the same way as the first. The evaluation has to be stable, not just accurate once.
02
Explainable output
A score with no reasoning cannot be used in a hiring decision. Every ranking carries the basis for it, so a person can agree or overrule it.
03
Messy document input
Scans, exports, tables, multi-column layouts, and files that are technically PDFs and practically images.
04
The human decision boundary
The product ranks and explains. It does not hire. Where the system stops is a design decision, not an omission.
05 / the principle
Keep people where judgement matters.
The temptation with a capable model is to let it decide.
Screening is a good candidate for automation because it is repetitive reading. Hiring is not, because it carries consequences a system cannot own.
So the product removes the reading work and hands back a ranked, explained shortlist.
The decision stays with the person accountable for it.
start here
Have an AI idea that needs to become a product?
The gap between a working prototype and a system people rely on is where most of the real work sits.
Discuss an AI product