AI · reading
The weakest document in hiring is the one everything else is built on.
A brief is three bullet points written between meetings, then read differently by everyone who touches it. A CV is a list of titles with the contribution left out. Both get treated as if they were specifications.
The reading layer is what closes the distance — and everything it produces carries the sentence it came from, so you are never asked to take its word for anything.
- Every value carries its source line
- Corrections hold in your workspace
- Gaps are named, not guessed
Extracted · with its source
- Notice period90 days
“I'd need to serve ninety days in full.”
- Current locationPune
“Based in Pune since 2021, open to relocating.”
- Team size owned6 engineers
“Ran the on-call rota for six engineers.”
Correct one and the correction stays corrected — your phrasing, your clients' job titles, in your workspace rather than in a global model.
The problem, specifically
Four holes in a brief that looks finished.
Below is a requirement of the kind that arrives every week. It is not a bad brief by the standards of the category — it is a completely normal one, which is the point. Four of its six lines carry a hole that the recruiter cannot see and the candidate pays for.
The brief, as written
Senior Engineering Manager
- Lead a team of 8–12 engineers across two squads
- Own delivery for the payments platform
- 8+ years of engineering experience, 3+ in a management role
- Experience with Kubernetes and Terraform
- Strong communication and stakeholder management skills
- Drive technical strategy and the quarterly roadmap
It records the duties, not the problem
Nothing says why the role exists now, what has been slipping without it, or what has to be different in a year. That is the part said out loud in the kickoff and never written down.
Nothing is ranked
Six requirements in one list, with no indication which two would sink the hire. A recruiter cannot tell a real requirement from a preference somebody added because it sounded useful.
It names the tool, not the capability
Written this way, somebody who did the same work on a different stack is filtered out at the search stage — and nobody ever finds out they were rejected.
Success is undefined
What the person must know is specified. What they must have achieved by month six is not, so every interviewer invents their own bar and the debrief becomes an argument.
What it establishes
Two understandings, built from different material.
Neither is a summary. A summary is shorter than the document; these are longer, because they contain what the document left out — and each is versioned, so you can see what changed and who confirmed it.
An understanding of the role
Built from the requirement and from the person who owns it. The business problem behind the role, what genuinely disqualifies somebody, and what good looks like six months in — the parts that were said out loud and never written down.
An understanding of the candidate
Built from the CV and from what the candidate said themselves. Not the words on the page but what the experience demonstrates, the context it happened in, and the depth of the contribution behind each line.
- It asks rather than assumes —
Where the role is silent, it works out what it does not know and puts the question to the person hiring — a conversation, not a longer form. The answers become part of the understanding.
- Confirmed means settled —
When a human confirms a section, it stops being second-guessed. The next pass leaves it alone, which is the difference between a tool that learns and one that keeps asking.
- Versioned, with the change visible —
An understanding is not overwritten. You can see what it used to say, what changed, and what the change was based on.
How reading fails elsewhere
Three failures that look like success on a demo document.
Every product in this category reads a clean, well-structured CV correctly. These are the three ways the same products fail on the documents that actually arrive, and each one is invisible until somebody audits a rejection.
Vocabulary transfer
Candidates and agencies paste the posting's language into a summary paragraph. A reader matching on overlap cannot tell that apart from somebody who has done the work — so the better copy-editor wins.
A candidate whose summary echoes the job description ranks above one whose project bullets prove they can do it.
Shape hidden by shared vocabulary
Adjacent specialists share terminology with a role without sharing its shape. On vocabulary alone they genuinely are equivalent.
A chief architect rated below a data-migration specialist, because the specialist happens to carry “enterprise architecture” in their title.
Seniority stripped out
Ten years owning a small workstream and ten years inside a large programme list the same technologies. Nothing in a keyword pass distinguishes them.
A nurse practitioner with fifteen years in intensive care, filtered out for writing “ICU” where the posting said “critical care unit”.
The guarantees
Three promises that survive contact with a bad document.
Extraction is easy to demonstrate and hard to keep. These are the parts that decide whether a recruiter is still using it in month three.
- A value without its source is not shown —
Every field carries the line it was read from. Confirming it takes a second and does not mean opening the document again — and a value whose source cannot be produced does not appear.
- Corrections belong to your workspace —
Fix an extraction once and the fix holds. Your phrasing, your clients' job titles, the shorthand your team uses. Nobody files a ticket and nothing is retrained somewhere else.
- The verdict opens per requirement —
What the role asked for, what the evidence shows, and the requirements nothing evidences at all — stated as gaps rather than folded into a number. A shortlist you can defend line by line is a different object from a ranked list you have to trust.
What you receive
Five things, not a score.
- A verdict per requirement —
Found by name, demonstrated by the work described, or not found — settled one requirement at a time, so a peripheral item never quietly steals attention from a central one.
- The resume text behind each one —
Anything other than “not found” carries the exact line that supports it. A positive verdict that cannot point at one is rejected before you see it.
- The gaps, stated as gaps —
The requirements nothing evidences at all, named rather than folded into a lower number.
- A boundary flag where it is genuinely close —
A candidate whose verdict would flip under a slightly different framing of the role is marked for a human read rather than pushed to one side of a line and hidden.
- A trail you can reconstruct —
Enough to explain any position on the list months later — to a hiring manager, to the candidate, or to somebody auditing the rejection.
What reads it next
An understanding is only worth building if something downstream uses it.
Matching
Searches the database against what the role actually needs rather than the keywords it happened to contain — and names the requirement that blocked each person.
Interview kits
The questions follow from what the role still has not established, so two interviewers stop covering the same ground.
Prep briefs
What to probe with this candidate on this role, drawn from the gap between the two understandings.
Feedback letters
A rejection that can name the actual gap, because the gap was recorded rather than inferred afterwards.
Objections first
What a recruiting lead asks in the second meeting.
›Our JDs are terrible. Does that break it?
That is the case it is built for. A thin brief is not rejected — it is the starting point, and what it does not say becomes the set of questions put to the person who owns the role.
›Who has to do the extra work?
The hiring manager, once, in a conversation rather than a form — and only about the things the brief left open. The alternative is the same conversation happening four times in four interviews.
›What if it reads something wrong?
You see the line it read. That is the whole design: being wrong is visible in about a second rather than discovered after three interviews on a misunderstanding.
›Do corrections leak between clients or tenants?
No. A correction applies in your workspace. It is not training anything that anybody else sees.
›Is this a score?
No, and deliberately. A verdict opens per requirement — what was asked for, what the evidence shows, what nothing evidences. The number that usually replaces all of that is the thing nobody can argue with, which is why it does not help.
Book a demo
Bring the worst brief you have.
The interesting document is the badly written one. Bring that, and a role you are actually struggling to fill.
- Give it a three-line JD and show me the questions it comes back with
- Open an extracted field and show me the sentence behind it
- Correct one, run another CV, and show me the correction held
- Open a verdict and show me a requirement nothing evidences