Resumes become records, not attachments

Raw extraction gives you the words on the page. Normalisation maps them to values you can filter on. React, ReactJS and React.js become one skill. Senior Software Engineer II resolves to a seniority band. A location becomes a place that can be measured for travel time rather than a string.

Skills, tools and certifications are held in separate fields rather than one blended list, because they answer different questions. AWS Solutions Architect is a certification with an expiry; AWS is a platform; Terraform is a tool. Collapsing them is why so many databases cannot answer a precise brief.

Seniority is the field that deserves the most suspicion, and the interface treats it that way. A title means different things in different companies and different markets, and a five-person startup and a bank do not use the word senior to describe the same job. The band is derived from title, scope and years together, it is always displayed as an inference rather than as a fact, and it is the field recruiters correct most often, which is exactly as it should be.

Parsed values are presented next to the source text they came from. Anything the parser inferred rather than read is marked as an inference, and low-confidence fields are highlighted rather than silently accepted.

This is not a formality. A parser that writes confidently wrong values into a database quietly poisons every search that runs afterwards, and nobody notices for months. A ten-second review at capture prevents the kind of error that is very expensive to find later.

The economics of that ten seconds are worth spelling out. A wrong seniority band on one record costs nothing until the day a search for mid-level candidates excludes exactly the right person, and at that point the cost is a placement nobody knows they lost. Errors in a candidate database are not noticed as errors; they are noticed as an absence, which is why they survive for years.

The uploaded document is kept alongside the parsed record, and every subsequent version is kept too. A two-year-old resume frequently holds a role the newest one has dropped, and clients periodically ask for the original rather than a formatted version.

A branded, formatted version can be generated for submittal, with contact details masked where the client relationship requires it. The original, the parsed text and the formatted output are three artefacts on one record, not three copies in three places.

Keeping the original also means a parse is never a one-way door. When the synonym list improves, or a title mapping is corrected, or a better extraction path is added for a format that used to fail, the stored document can be parsed again. A system that discards the source after extraction has permanently capped the quality of its own records at whatever the parser could do on the day the file arrived.

A folder of resumes, an inbox of applications or an export from a previous system can be processed as a batch. Each row reports its own outcome: parsed cleanly, parsed with low confidence, matched to an existing candidate, or failed with a reason.

Batches are reviewable before they are committed, and duplicate detection runs inside the batch as well as against the existing database, so importing nine hundred resumes does not create four hundred duplicates.

In-batch checking is the part that is easy to leave out and expensive to omit. A folder exported from an older system routinely contains the same person three times, under three filenames, from three different years. Checking each row only against the existing database would let all three through as new records, and the clean-up afterwards costs more than the import saved.

Parsing handles the common English-language resume formats well, including the dense multi-page format usual in Indian IT staffing and the one-page format usual in the United States. Non-English resumes are extracted and stored but structured with lower confidence, and the record says so rather than pretending otherwise.

Where a document cannot be parsed at all, it is stored, flagged and queued for manual entry rather than dropped. A silent failure is worse than a visible one.

The formats that reliably cause trouble are worth naming so you can test them rather than discover them. Resumes exported from design tools as a single image, heavily tabular consultancy CVs where a project grid carries the actual experience, federal-style documents that run to fifteen pages, and any layout where the dates sit in a narrow column beside the employer. All of those are readable; all of them are more likely to arrive with fields flagged for review.

The first real test of a parser is never the demo file, it is the backlog: a folder of several thousand documents accumulated over years, in every format anybody ever sent. The mistake is to budget for the runtime and not for the review, because the batch runs overnight and the queue of low-confidence rows is what actually takes the week.

The order that saves the most time is to parse a sample of fifty first, deliberately picked across your worst formats rather than your cleanest, and read what came out. Almost every desk finds two or three synonym problems specific to its market, where an industry term or an internal job title maps to the wrong skill or the wrong band. Promoting those corrections to the tenant override list before the full batch runs applies them to every subsequent document instead of to fifty records by hand.

Then run the batch and work the outcomes by category rather than by row. Everything that failed extraction goes to manual entry. Everything that matched an existing candidate goes to the duplicate queue. Everything flagged low-confidence gets reviewed for the two or three fields that were actually uncertain, not re-read end to end. The credit cost for the batch is shown before it runs, with scanned pages weighted higher, so the size of the job is a decision rather than a surprise.

It cannot recover what the document never contained. Reason for leaving, notice period, salary expectation, right to work, willingness to relocate and whether the person is actually looking are the fields that decide whether a candidate is submittable, and almost none of them appear on a resume. Those come from a conversation, and a parsed record is the thing that makes the conversation shorter, not the thing that replaces it.

It also cannot judge truth. A resume is a claim, and the parser reads claims exactly as written. Overstated seniority, a contract described as a permanent role, dates rounded to hide a gap: all of that is structured faithfully and none of it is verified. Screening, referencing and verification are separate pieces of work, and a confidently structured record can make an unverified claim look more solid than it did as prose, which is worth being alert to.

And it is bounded by the document quality you actually receive. Non-English resumes and unusual layouts are extracted and stored at lower confidence, and the record says so. A document that cannot be parsed is flagged and queued rather than dropped. We do not publish a headline accuracy number, because one measured on clean single-column English resumes would describe our sample rather than your inbox, and it would be the most misleading number on this page.